Remove comment

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: Jamison Lahman <jamison@lahman.dev>

.git-blame-ignore-revs

@@ -0,0 +1,8 @@
+# Exclude these commits from git blame (e.g. mass reformatting).
+# These are ignored by GitHub automatically.
+# To enable this locally, run:
+#
+#    git config blame.ignoreRevsFile .git-blame-ignore-revs
+
+3134e5f840c12c8f32613ce520101a047c89dcc2  # refactor(whitespace): rm temporary react fragments (#7161)
+ed3f72bc75f3e3a9ae9e4d8cd38278f9c97e78b4  # refactor(whitespace): rm react fragment #7190

.github/CODEOWNERS

@@ -1,3 +1,10 @@
 * @onyx-dot-app/onyx-core-team
 # Helm charts Owners
 /helm/ @justin-tahara
+
+# Web standards updates
+/web/STANDARDS.md @raunakab @Weves
+
+# Agent context files
+/CLAUDE.md.template @Weves
+/AGENTS.md.template @Weves

.github/actions/setup-python-and-install-dependencies/action.yml

@@ -7,12 +7,6 @@ inputs:
 runs:
   using: "composite"
   steps:
-    - name: Setup uv
-      uses: astral-sh/setup-uv@caf0cab7a618c569241d31dcd442f54681755d39 # ratchet:astral-sh/setup-uv@v3
-      # TODO: Enable caching once there is a uv.lock file checked in.
-      # with:
-      #   enable-cache: true
-
     - name: Compute requirements hash
       id: req-hash
       shell: bash
@@ -28,6 +22,8 @@ runs:
         done <<< "$REQUIREMENTS"
         echo "hash=$(echo "$hash" | sha256sum | cut -d' ' -f1)" >> "$GITHUB_OUTPUT"
 
+    # NOTE: This comes before Setup uv since clean-ups run in reverse chronological order
+    # such that Setup uv's prune-cache is able to prune the cache before we upload.
     - name: Cache uv cache directory
       uses: runs-on/cache@50350ad4242587b6c8c2baa2e740b1bc11285ff4 # ratchet:runs-on/cache@v4
       with:
@@ -36,6 +32,14 @@ runs:
         restore-keys: |
           ${{ runner.os }}-uv-
 
+    - name: Setup uv
+      uses: astral-sh/setup-uv@ed21f2f24f8dd64503750218de024bcf64c7250a # ratchet:astral-sh/setup-uv@v7
+      with:
+        version: "0.9.9"
+      # TODO: Enable caching once there is a uv.lock file checked in.
+      # with:
+      #   enable-cache: true
+
     - name: Setup Python
       uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # ratchet:actions/setup-python@v5
       with:

.github/pull_request_template.md

@@ -1,10 +1,10 @@
 ## Description
 
-[Provide a brief description of the changes in this PR]
+<!--- Provide a brief description of the changes in this PR --->
 
 ## How Has This Been Tested?
 
-[Describe the tests you ran to verify your changes]
+<!--- Describe the tests you ran to verify your changes --->
 
 ## Additional Options
 

.github/workflows/deployment.yml

@@ -11,7 +11,6 @@ on:
 permissions: {}
 
 env:
-  IS_DRY_RUN: ${{ github.event_name == 'workflow_dispatch' }}
   EDGE_TAG: ${{ startsWith(github.ref_name, 'nightly-latest') }}
 
 jobs:
@@ -31,27 +30,46 @@ jobs:
       is-beta: ${{ steps.check.outputs.is-beta }}
       is-stable-standalone: ${{ steps.check.outputs.is-stable-standalone }}
       is-beta-standalone: ${{ steps.check.outputs.is-beta-standalone }}
+      is-test-run: ${{ steps.check.outputs.is-test-run }}
       sanitized-tag: ${{ steps.check.outputs.sanitized-tag }}
       short-sha: ${{ steps.check.outputs.short-sha }}
     steps:
       - name: Check which components to build and version info
         id: check
+        env:
+          EVENT_NAME: ${{ github.event_name }}
         run: |
           TAG="${GITHUB_REF_NAME}"
           # Sanitize tag name by replacing slashes with hyphens (for Docker tag compatibility)
           SANITIZED_TAG=$(echo "$TAG" | tr '/' '-')
+          SHORT_SHA="${GITHUB_SHA::7}"
+
+          # Initialize all flags to false
           IS_CLOUD=false
+          IS_NIGHTLY=false
+          IS_VERSION_TAG=false
+          IS_STABLE=false
+          IS_BETA=false
+          IS_STABLE_STANDALONE=false
+          IS_BETA_STANDALONE=false
+          IS_PROD_TAG=false
+          IS_TEST_RUN=false
           BUILD_DESKTOP=false
           BUILD_WEB=false
           BUILD_WEB_CLOUD=false
           BUILD_BACKEND=true
           BUILD_MODEL_SERVER=true
-          IS_STABLE=false
-          IS_BETA=false
-          IS_STABLE_STANDALONE=false
-          IS_BETA_STANDALONE=false
 
-          # Version checks (for web - any stable version)
+          # Determine tag type based on pattern matching (do regex checks once)
+          if [[ "$TAG" == *cloud* ]]; then
+            IS_CLOUD=true
+          fi
+          if [[ "$TAG" == nightly* ]]; then
+            IS_NIGHTLY=true
+          fi
+          if [[ "$TAG" =~ ^v[0-9]+\.[0-9]+\.[0-9]+ ]]; then
+            IS_VERSION_TAG=true
+          fi
           if [[ "$TAG" =~ ^v[0-9]+\.[0-9]+\.[0-9]+$ ]]; then
             IS_STABLE=true
           fi
@@ -59,26 +77,35 @@ jobs:
             IS_BETA=true
           fi
 
-          if [[ "$TAG" == *cloud* ]]; then
-            IS_CLOUD=true
+          # Determine what to build based on tag type
+          if [[ "$IS_CLOUD" == "true" ]]; then
             BUILD_WEB_CLOUD=true
           else
             BUILD_WEB=true
-            # Skip desktop builds on beta tags
-            if [[ "$IS_BETA" != "true" ]]; then
+            # Skip desktop builds on beta tags and nightly runs
+            if [[ "$IS_BETA" != "true" ]] && [[ "$IS_NIGHTLY" != "true" ]]; then
               BUILD_DESKTOP=true
             fi
           fi
 
-          # Version checks (for backend/model-server - stable version excluding cloud tags)
-          if [[ "$TAG" =~ ^v[0-9]+\.[0-9]+\.[0-9]+$ ]] && [[ "$TAG" != *cloud* ]]; then
+          # Standalone version checks (for backend/model-server - version excluding cloud tags)
+          if [[ "$IS_STABLE" == "true" ]] && [[ "$IS_CLOUD" != "true" ]]; then
             IS_STABLE_STANDALONE=true
           fi
-          if [[ "$TAG" =~ ^v[0-9]+\.[0-9]+\.[0-9]+-beta\.[0-9]+$ ]] && [[ "$TAG" != *cloud* ]]; then
+          if [[ "$IS_BETA" == "true" ]] && [[ "$IS_CLOUD" != "true" ]]; then
             IS_BETA_STANDALONE=true
           fi
 
-          SHORT_SHA="${GITHUB_SHA::7}"
+          # Determine if this is a production tag
+          # Production tags are: version tags (v1.2.3*) or nightly tags
+          if [[ "$IS_VERSION_TAG" == "true" ]] || [[ "$IS_NIGHTLY" == "true" ]]; then
+            IS_PROD_TAG=true
+          fi
+
+          # Determine if this is a test run (workflow_dispatch on non-production ref)
+          if [[ "$EVENT_NAME" == "workflow_dispatch" ]] && [[ "$IS_PROD_TAG" != "true" ]]; then
+            IS_TEST_RUN=true
+          fi
           {
             echo "build-desktop=$BUILD_DESKTOP"
             echo "build-web=$BUILD_WEB"
@@ -90,6 +117,7 @@ jobs:
             echo "is-beta=$IS_BETA"
             echo "is-stable-standalone=$IS_STABLE_STANDALONE"
             echo "is-beta-standalone=$IS_BETA_STANDALONE"
+            echo "is-test-run=$IS_TEST_RUN"
             echo "sanitized-tag=$SANITIZED_TAG"
             echo "short-sha=$SHORT_SHA"
           } >> "$GITHUB_OUTPUT"
@@ -106,8 +134,9 @@ jobs:
           fetch-depth: 0
 
       - name: Setup uv
-        uses: astral-sh/setup-uv@ed21f2f24f8dd64503750218de024bcf64c7250a # ratchet:astral-sh/setup-uv@v7.1.5
+        uses: astral-sh/setup-uv@ed21f2f24f8dd64503750218de024bcf64c7250a # ratchet:astral-sh/setup-uv@v7
         with:
+          version: "0.9.9"
           # NOTE: This isn't caching much and zizmor suggests this could be poisoned, so disable.
           enable-cache: false
 
@@ -145,14 +174,14 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - platform: 'macos-latest' # Build a universal image for macOS.
-            args: '--target universal-apple-darwin'
-          - platform: 'ubuntu-24.04'
-            args: '--bundles deb,rpm'
-          - platform: 'ubuntu-24.04-arm' # Only available in public repos.
-            args: '--bundles deb,rpm'
-          - platform: 'windows-latest'
-            args: ''
+          - platform: "macos-latest" # Build a universal image for macOS.
+            args: "--target universal-apple-darwin"
+          - platform: "ubuntu-24.04"
+            args: "--bundles deb,rpm"
+          - platform: "ubuntu-24.04-arm" # Only available in public repos.
+            args: "--bundles deb,rpm"
+          - platform: "windows-latest"
+            args: ""
 
     runs-on: ${{ matrix.platform }}
     timeout-minutes: 90
@@ -200,9 +229,9 @@ jobs:
         working-directory: ./desktop
         env:
           SHORT_SHA: ${{ needs.determine-builds.outputs.short-sha }}
-          EVENT_NAME: ${{ github.event_name }}
+          IS_TEST_RUN: ${{ needs.determine-builds.outputs.is-test-run }}
         run: |
-          if [ "${EVENT_NAME}" == "workflow_dispatch" ]; then
+          if [ "${IS_TEST_RUN}" == "true" ]; then
             VERSION="0.0.0-dev+${SHORT_SHA}"
           else
             VERSION="${GITHUB_REF_NAME#v}"
@@ -227,9 +256,11 @@ jobs:
         if: runner.os == 'Windows'
         working-directory: ./desktop
         shell: pwsh
+        env:
+          IS_TEST_RUN: ${{ needs.determine-builds.outputs.is-test-run }}
         run: |
           # Windows MSI requires numeric-only build metadata, so we skip the SHA suffix
-          if ("${{ github.event_name }}" -eq "workflow_dispatch") {
+          if ($env:IS_TEST_RUN -eq "true") {
             $VERSION = "0.0.0"
           } else {
             # Strip 'v' prefix and any pre-release suffix (e.g., -beta.13) for MSI compatibility
@@ -258,9 +289,9 @@ jobs:
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         with:
-          tagName: ${{ github.event_name != 'workflow_dispatch' && 'v__VERSION__' || format('v0.0.0-dev+{0}', needs.determine-builds.outputs.short-sha) }}
-          releaseName: ${{ github.event_name != 'workflow_dispatch' && 'v__VERSION__' || format('v0.0.0-dev+{0}', needs.determine-builds.outputs.short-sha) }}
-          releaseBody: 'See the assets to download this version and install.'
+          tagName: ${{ needs.determine-builds.outputs.is-test-run != 'true' && 'v__VERSION__' || format('v0.0.0-dev+{0}', needs.determine-builds.outputs.short-sha) }}
+          releaseName: ${{ needs.determine-builds.outputs.is-test-run != 'true' && 'v__VERSION__' || format('v0.0.0-dev+{0}', needs.determine-builds.outputs.short-sha) }}
+          releaseBody: "See the assets to download this version and install."
           releaseDraft: true
           prerelease: false
           args: ${{ matrix.args }}
@@ -290,7 +321,7 @@ jobs:
         id: meta
         uses: docker/metadata-action@c299e40c65443455700f0fdfc63efafe5b349051 # ratchet:docker/metadata-action@v5
         with:
-          images: ${{ github.event_name == 'workflow_dispatch' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
+          images: ${{ needs.determine-builds.outputs.is-test-run == 'true' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
           flavor: |
             latest=false
 
@@ -320,7 +351,7 @@ jobs:
           cache-to: |
             type=inline
             type=registry,ref=${{ env.RUNS_ON_ECR_CACHE }}:web-cache-amd64,mode=max
-          outputs: type=image,name=${{ github.event_name == 'workflow_dispatch' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }},push-by-digest=true,name-canonical=true,push=true
+          outputs: type=image,name=${{ needs.determine-builds.outputs.is-test-run == 'true' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }},push-by-digest=true,name-canonical=true,push=true
           no-cache: ${{ vars.DOCKER_NO_CACHE == 'true' }}
 
   build-web-arm64:
@@ -348,7 +379,7 @@ jobs:
         id: meta
         uses: docker/metadata-action@c299e40c65443455700f0fdfc63efafe5b349051 # ratchet:docker/metadata-action@v5
         with:
-          images: ${{ github.event_name == 'workflow_dispatch' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
+          images: ${{ needs.determine-builds.outputs.is-test-run == 'true' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
           flavor: |
             latest=false
 
@@ -378,7 +409,7 @@ jobs:
           cache-to: |
             type=inline
             type=registry,ref=${{ env.RUNS_ON_ECR_CACHE }}:web-cache-arm64,mode=max
-          outputs: type=image,name=${{ github.event_name == 'workflow_dispatch' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }},push-by-digest=true,name-canonical=true,push=true
+          outputs: type=image,name=${{ needs.determine-builds.outputs.is-test-run == 'true' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }},push-by-digest=true,name-canonical=true,push=true
           no-cache: ${{ vars.DOCKER_NO_CACHE == 'true' }}
 
   merge-web:
@@ -410,18 +441,18 @@ jobs:
         id: meta
         uses: docker/metadata-action@c299e40c65443455700f0fdfc63efafe5b349051 # ratchet:docker/metadata-action@v5
         with:
-          images: ${{ github.event_name == 'workflow_dispatch' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
+          images: ${{ needs.determine-builds.outputs.is-test-run == 'true' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
           flavor: |
             latest=false
           tags: |
-            type=raw,value=${{ github.event_name == 'workflow_dispatch' && format('web-{0}', needs.determine-builds.outputs.sanitized-tag) || github.ref_name }}
-            type=raw,value=${{ github.event_name != 'workflow_dispatch' && needs.determine-builds.outputs.is-stable == 'true' && 'latest' || '' }}
-            type=raw,value=${{ github.event_name != 'workflow_dispatch' && env.EDGE_TAG == 'true' && 'edge' || '' }}
-            type=raw,value=${{ github.event_name != 'workflow_dispatch' && needs.determine-builds.outputs.is-beta == 'true' && 'beta' || '' }}
+            type=raw,value=${{ needs.determine-builds.outputs.is-test-run == 'true' && format('web-{0}', needs.determine-builds.outputs.sanitized-tag) || github.ref_name }}
+            type=raw,value=${{ needs.determine-builds.outputs.is-test-run != 'true' && needs.determine-builds.outputs.is-stable == 'true' && 'latest' || '' }}
+            type=raw,value=${{ needs.determine-builds.outputs.is-test-run != 'true' && env.EDGE_TAG == 'true' && 'edge' || '' }}
+            type=raw,value=${{ needs.determine-builds.outputs.is-test-run != 'true' && needs.determine-builds.outputs.is-beta == 'true' && 'beta' || '' }}
 
       - name: Create and push manifest
         env:
-          IMAGE_REPO: ${{ github.event_name == 'workflow_dispatch' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
+          IMAGE_REPO: ${{ needs.determine-builds.outputs.is-test-run == 'true' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
           AMD64_DIGEST: ${{ needs.build-web-amd64.outputs.digest }}
           ARM64_DIGEST: ${{ needs.build-web-arm64.outputs.digest }}
           META_TAGS: ${{ steps.meta.outputs.tags }}
@@ -456,7 +487,7 @@ jobs:
         id: meta
         uses: docker/metadata-action@c299e40c65443455700f0fdfc63efafe5b349051 # ratchet:docker/metadata-action@v5
         with:
-          images: ${{ github.event_name == 'workflow_dispatch' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
+          images: ${{ needs.determine-builds.outputs.is-test-run == 'true' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
           flavor: |
             latest=false
 
@@ -494,7 +525,7 @@ jobs:
           cache-to: |
             type=inline
             type=registry,ref=${{ env.RUNS_ON_ECR_CACHE }}:cloudweb-cache-amd64,mode=max
-          outputs: type=image,name=${{ github.event_name == 'workflow_dispatch' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }},push-by-digest=true,name-canonical=true,push=true
+          outputs: type=image,name=${{ needs.determine-builds.outputs.is-test-run == 'true' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }},push-by-digest=true,name-canonical=true,push=true
           no-cache: ${{ vars.DOCKER_NO_CACHE == 'true' }}
 
   build-web-cloud-arm64:
@@ -522,7 +553,7 @@ jobs:
         id: meta
         uses: docker/metadata-action@c299e40c65443455700f0fdfc63efafe5b349051 # ratchet:docker/metadata-action@v5
         with:
-          images: ${{ github.event_name == 'workflow_dispatch' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
+          images: ${{ needs.determine-builds.outputs.is-test-run == 'true' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
           flavor: |
             latest=false
 
@@ -560,7 +591,7 @@ jobs:
           cache-to: |
             type=inline
             type=registry,ref=${{ env.RUNS_ON_ECR_CACHE }}:cloudweb-cache-arm64,mode=max
-          outputs: type=image,name=${{ github.event_name == 'workflow_dispatch' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }},push-by-digest=true,name-canonical=true,push=true
+          outputs: type=image,name=${{ needs.determine-builds.outputs.is-test-run == 'true' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }},push-by-digest=true,name-canonical=true,push=true
           no-cache: ${{ vars.DOCKER_NO_CACHE == 'true' }}
 
   merge-web-cloud:
@@ -592,15 +623,15 @@ jobs:
         id: meta
         uses: docker/metadata-action@c299e40c65443455700f0fdfc63efafe5b349051 # ratchet:docker/metadata-action@v5
         with:
-          images: ${{ github.event_name == 'workflow_dispatch' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
+          images: ${{ needs.determine-builds.outputs.is-test-run == 'true' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
           flavor: |
             latest=false
           tags: |
-            type=raw,value=${{ github.event_name == 'workflow_dispatch' && format('web-cloud-{0}', needs.determine-builds.outputs.sanitized-tag) || github.ref_name }}
+            type=raw,value=${{ needs.determine-builds.outputs.is-test-run == 'true' && format('web-cloud-{0}', needs.determine-builds.outputs.sanitized-tag) || github.ref_name }}
 
       - name: Create and push manifest
         env:
-          IMAGE_REPO: ${{ github.event_name == 'workflow_dispatch' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
+          IMAGE_REPO: ${{ needs.determine-builds.outputs.is-test-run == 'true' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
           AMD64_DIGEST: ${{ needs.build-web-cloud-amd64.outputs.digest }}
           ARM64_DIGEST: ${{ needs.build-web-cloud-arm64.outputs.digest }}
           META_TAGS: ${{ steps.meta.outputs.tags }}
@@ -635,7 +666,7 @@ jobs:
         id: meta
         uses: docker/metadata-action@c299e40c65443455700f0fdfc63efafe5b349051 # ratchet:docker/metadata-action@v5
         with:
-          images: ${{ github.event_name == 'workflow_dispatch' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
+          images: ${{ needs.determine-builds.outputs.is-test-run == 'true' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
           flavor: |
             latest=false
 
@@ -664,7 +695,7 @@ jobs:
           cache-to: |
             type=inline
             type=registry,ref=${{ env.RUNS_ON_ECR_CACHE }}:backend-cache-amd64,mode=max
-          outputs: type=image,name=${{ github.event_name == 'workflow_dispatch' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }},push-by-digest=true,name-canonical=true,push=true
+          outputs: type=image,name=${{ needs.determine-builds.outputs.is-test-run == 'true' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }},push-by-digest=true,name-canonical=true,push=true
           no-cache: ${{ vars.DOCKER_NO_CACHE == 'true' }}
 
   build-backend-arm64:
@@ -692,7 +723,7 @@ jobs:
         id: meta
         uses: docker/metadata-action@c299e40c65443455700f0fdfc63efafe5b349051 # ratchet:docker/metadata-action@v5
         with:
-          images: ${{ github.event_name == 'workflow_dispatch' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
+          images: ${{ needs.determine-builds.outputs.is-test-run == 'true' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
           flavor: |
             latest=false
 
@@ -721,7 +752,7 @@ jobs:
           cache-to: |
             type=inline
             type=registry,ref=${{ env.RUNS_ON_ECR_CACHE }}:backend-cache-arm64,mode=max
-          outputs: type=image,name=${{ github.event_name == 'workflow_dispatch' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }},push-by-digest=true,name-canonical=true,push=true
+          outputs: type=image,name=${{ needs.determine-builds.outputs.is-test-run == 'true' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }},push-by-digest=true,name-canonical=true,push=true
           no-cache: ${{ vars.DOCKER_NO_CACHE == 'true' }}
 
   merge-backend:
@@ -753,18 +784,18 @@ jobs:
         id: meta
         uses: docker/metadata-action@c299e40c65443455700f0fdfc63efafe5b349051 # ratchet:docker/metadata-action@v5
         with:
-          images: ${{ github.event_name == 'workflow_dispatch' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
+          images: ${{ needs.determine-builds.outputs.is-test-run == 'true' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
           flavor: |
             latest=false
           tags: |
-            type=raw,value=${{ github.event_name == 'workflow_dispatch' && format('backend-{0}', needs.determine-builds.outputs.sanitized-tag) || github.ref_name }}
-            type=raw,value=${{ github.event_name != 'workflow_dispatch' && needs.determine-builds.outputs.is-stable-standalone == 'true' && 'latest' || '' }}
-            type=raw,value=${{ github.event_name != 'workflow_dispatch' && env.EDGE_TAG == 'true' && 'edge' || '' }}
-            type=raw,value=${{ github.event_name != 'workflow_dispatch' && needs.determine-builds.outputs.is-beta-standalone == 'true' && 'beta' || '' }}
+            type=raw,value=${{ needs.determine-builds.outputs.is-test-run == 'true' && format('backend-{0}', needs.determine-builds.outputs.sanitized-tag) || github.ref_name }}
+            type=raw,value=${{ needs.determine-builds.outputs.is-test-run != 'true' && needs.determine-builds.outputs.is-stable-standalone == 'true' && 'latest' || '' }}
+            type=raw,value=${{ needs.determine-builds.outputs.is-test-run != 'true' && env.EDGE_TAG == 'true' && 'edge' || '' }}
+            type=raw,value=${{ needs.determine-builds.outputs.is-test-run != 'true' && needs.determine-builds.outputs.is-beta-standalone == 'true' && 'beta' || '' }}
 
       - name: Create and push manifest
         env:
-          IMAGE_REPO: ${{ github.event_name == 'workflow_dispatch' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
+          IMAGE_REPO: ${{ needs.determine-builds.outputs.is-test-run == 'true' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
           AMD64_DIGEST: ${{ needs.build-backend-amd64.outputs.digest }}
           ARM64_DIGEST: ${{ needs.build-backend-arm64.outputs.digest }}
           META_TAGS: ${{ steps.meta.outputs.tags }}
@@ -800,7 +831,7 @@ jobs:
         id: meta
         uses: docker/metadata-action@c299e40c65443455700f0fdfc63efafe5b349051 # ratchet:docker/metadata-action@v5
         with:
-          images: ${{ github.event_name == 'workflow_dispatch' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
+          images: ${{ needs.determine-builds.outputs.is-test-run == 'true' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
           flavor: |
             latest=false
 
@@ -833,7 +864,7 @@ jobs:
           cache-to: |
             type=inline
             type=registry,ref=${{ env.RUNS_ON_ECR_CACHE }}:model-server-cache-amd64,mode=max
-          outputs: type=image,name=${{ github.event_name == 'workflow_dispatch' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }},push-by-digest=true,name-canonical=true,push=true
+          outputs: type=image,name=${{ needs.determine-builds.outputs.is-test-run == 'true' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }},push-by-digest=true,name-canonical=true,push=true
           no-cache: ${{ env.EDGE_TAG != 'true' && vars.MODEL_SERVER_NO_CACHE == 'true' }}
           provenance: false
           sbom: false
@@ -864,7 +895,7 @@ jobs:
         id: meta
         uses: docker/metadata-action@c299e40c65443455700f0fdfc63efafe5b349051 # ratchet:docker/metadata-action@v5
         with:
-          images: ${{ github.event_name == 'workflow_dispatch' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
+          images: ${{ needs.determine-builds.outputs.is-test-run == 'true' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
           flavor: |
             latest=false
 
@@ -897,7 +928,7 @@ jobs:
           cache-to: |
             type=inline
             type=registry,ref=${{ env.RUNS_ON_ECR_CACHE }}:model-server-cache-arm64,mode=max
-          outputs: type=image,name=${{ github.event_name == 'workflow_dispatch' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }},push-by-digest=true,name-canonical=true,push=true
+          outputs: type=image,name=${{ needs.determine-builds.outputs.is-test-run == 'true' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }},push-by-digest=true,name-canonical=true,push=true
           no-cache: ${{ env.EDGE_TAG != 'true' && vars.MODEL_SERVER_NO_CACHE == 'true' }}
           provenance: false
           sbom: false
@@ -931,18 +962,18 @@ jobs:
         id: meta
         uses: docker/metadata-action@c299e40c65443455700f0fdfc63efafe5b349051 # ratchet:docker/metadata-action@v5
         with:
-          images: ${{ github.event_name == 'workflow_dispatch' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
+          images: ${{ needs.determine-builds.outputs.is-test-run == 'true' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
           flavor: |
             latest=false
           tags: |
-            type=raw,value=${{ github.event_name == 'workflow_dispatch' && format('model-server-{0}', needs.determine-builds.outputs.sanitized-tag) || github.ref_name }}
-            type=raw,value=${{ github.event_name != 'workflow_dispatch' && needs.determine-builds.outputs.is-stable-standalone == 'true' && 'latest' || '' }}
-            type=raw,value=${{ github.event_name != 'workflow_dispatch' && env.EDGE_TAG == 'true' && 'edge' || '' }}
-            type=raw,value=${{ github.event_name != 'workflow_dispatch' && needs.determine-builds.outputs.is-beta-standalone == 'true' && 'beta' || '' }}
+            type=raw,value=${{ needs.determine-builds.outputs.is-test-run == 'true' && format('model-server-{0}', needs.determine-builds.outputs.sanitized-tag) || github.ref_name }}
+            type=raw,value=${{ needs.determine-builds.outputs.is-test-run != 'true' && needs.determine-builds.outputs.is-stable-standalone == 'true' && 'latest' || '' }}
+            type=raw,value=${{ needs.determine-builds.outputs.is-test-run != 'true' && env.EDGE_TAG == 'true' && 'edge' || '' }}
+            type=raw,value=${{ needs.determine-builds.outputs.is-test-run != 'true' && needs.determine-builds.outputs.is-beta-standalone == 'true' && 'beta' || '' }}
 
       - name: Create and push manifest
         env:
-          IMAGE_REPO: ${{ github.event_name == 'workflow_dispatch' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
+          IMAGE_REPO: ${{ needs.determine-builds.outputs.is-test-run == 'true' && env.RUNS_ON_ECR_CACHE || env.REGISTRY_IMAGE }}
           AMD64_DIGEST: ${{ needs.build-model-server-amd64.outputs.digest }}
           ARM64_DIGEST: ${{ needs.build-model-server-arm64.outputs.digest }}
           META_TAGS: ${{ steps.meta.outputs.tags }}
@@ -975,7 +1006,7 @@ jobs:
           max_attempts: 3
           retry_wait_seconds: 10
           command: |
-            if [ "${{ github.event_name }}" == "workflow_dispatch" ]; then
+            if [ "${{ needs.determine-builds.outputs.is-test-run }}" == "true" ]; then
               SCAN_IMAGE="${{ env.RUNS_ON_ECR_CACHE }}:web-${{ needs.determine-builds.outputs.sanitized-tag }}"
             else
               SCAN_IMAGE="docker.io/${{ env.REGISTRY_IMAGE }}:${{ github.ref_name }}"
@@ -1015,7 +1046,7 @@ jobs:
           max_attempts: 3
           retry_wait_seconds: 10
           command: |
-            if [ "${{ github.event_name }}" == "workflow_dispatch" ]; then
+            if [ "${{ needs.determine-builds.outputs.is-test-run }}" == "true" ]; then
               SCAN_IMAGE="${{ env.RUNS_ON_ECR_CACHE }}:web-cloud-${{ needs.determine-builds.outputs.sanitized-tag }}"
             else
               SCAN_IMAGE="docker.io/${{ env.REGISTRY_IMAGE }}:${{ github.ref_name }}"
@@ -1060,7 +1091,7 @@ jobs:
           max_attempts: 3
           retry_wait_seconds: 10
           command: |
-            if [ "${{ github.event_name }}" == "workflow_dispatch" ]; then
+            if [ "${{ needs.determine-builds.outputs.is-test-run }}" == "true" ]; then
               SCAN_IMAGE="${{ env.RUNS_ON_ECR_CACHE }}:backend-${{ needs.determine-builds.outputs.sanitized-tag }}"
             else
               SCAN_IMAGE="docker.io/${{ env.REGISTRY_IMAGE }}:${{ github.ref_name }}"
@@ -1102,7 +1133,7 @@ jobs:
           max_attempts: 3
           retry_wait_seconds: 10
           command: |
-            if [ "${{ github.event_name }}" == "workflow_dispatch" ]; then
+            if [ "${{ needs.determine-builds.outputs.is-test-run }}" == "true" ]; then
               SCAN_IMAGE="${{ env.RUNS_ON_ECR_CACHE }}:model-server-${{ needs.determine-builds.outputs.sanitized-tag }}"
             else
               SCAN_IMAGE="docker.io/${{ env.REGISTRY_IMAGE }}:${{ github.ref_name }}"
@@ -1121,6 +1152,7 @@ jobs:
 
   notify-slack-on-failure:
     needs:
+      - determine-builds
       - build-desktop
       - build-web-amd64
       - build-web-arm64
@@ -1134,7 +1166,7 @@ jobs:
       - build-model-server-amd64
       - build-model-server-arm64
       - merge-model-server
-    if: always() && (needs.build-desktop.result == 'failure' || needs.build-web-amd64.result == 'failure' || needs.build-web-arm64.result == 'failure' || needs.merge-web.result == 'failure' || needs.build-web-cloud-amd64.result == 'failure' || needs.build-web-cloud-arm64.result == 'failure' || needs.merge-web-cloud.result == 'failure' || needs.build-backend-amd64.result == 'failure' || needs.build-backend-arm64.result == 'failure' || needs.merge-backend.result == 'failure' || needs.build-model-server-amd64.result == 'failure' || needs.build-model-server-arm64.result == 'failure' || needs.merge-model-server.result == 'failure') && github.event_name != 'workflow_dispatch'
+    if: always() && (needs.build-desktop.result == 'failure' || needs.build-web-amd64.result == 'failure' || needs.build-web-arm64.result == 'failure' || needs.merge-web.result == 'failure' || needs.build-web-cloud-amd64.result == 'failure' || needs.build-web-cloud-arm64.result == 'failure' || needs.merge-web-cloud.result == 'failure' || needs.build-backend-amd64.result == 'failure' || needs.build-backend-arm64.result == 'failure' || needs.merge-backend.result == 'failure' || needs.build-model-server-amd64.result == 'failure' || needs.build-model-server-arm64.result == 'failure' || needs.merge-model-server.result == 'failure') && needs.determine-builds.outputs.is-test-run != 'true'
     # NOTE: Github-hosted runners have about 20s faster queue times and are preferred here.
     runs-on: ubuntu-slim
     timeout-minutes: 90

.github/workflows/merge-group.yml

@@ -0,0 +1,31 @@
+name: Merge Group-Specific
+
+on:
+  merge_group:
+
+permissions:
+  contents: read
+
+jobs:
+  # This job immediately succeeds to satisfy branch protection rules on merge_group events.
+  # There is a similarly named "required" job in pr-integration-tests.yml which runs the actual
+  # integration tests. That job runs on both pull_request and merge_group events, and this job
+  # exists solely to provide a fast-passing check with the same name for branch protection.
+  # The actual tests remain enforced on presubmit (pull_request events).
+  required:
+    runs-on: ubuntu-latest
+    timeout-minutes: 45
+    steps:
+      - name: Success
+        run: echo "Success"
+  # This job immediately succeeds to satisfy branch protection rules on merge_group events.
+  # There is a similarly named "playwright-required" job in pr-playwright-tests.yml which runs
+  # the actual playwright tests. That job runs on both pull_request and merge_group events, and
+  # this job exists solely to provide a fast-passing check with the same name for branch protection.
+  # The actual tests remain enforced on presubmit (pull_request events).
+  playwright-required:
+    runs-on: ubuntu-latest
+    timeout-minutes: 45
+    steps:
+      - name: Success
+        run: echo "Success"

.github/workflows/pr-database-tests.yml

@@ -0,0 +1,62 @@
+name: Database Tests
+concurrency:
+  group: Database-Tests-${{ github.workflow }}-${{ github.head_ref || github.event.workflow_run.head_branch || github.run_id }}
+  cancel-in-progress: true
+
+on:
+  merge_group:
+  pull_request:
+    branches:
+      - main
+      - "release/**"
+  push:
+    tags:
+      - "v*.*.*"
+
+permissions:
+  contents: read
+
+jobs:
+  database-tests:
+    runs-on:
+      - runs-on
+      - runner=2cpu-linux-arm64
+      - "run-id=${{ github.run_id }}-database-tests"
+    timeout-minutes: 45
+    steps:
+      - uses: runs-on/action@cd2b598b0515d39d78c38a02d529db87d2196d1e # ratchet:runs-on/action@v2
+
+      - name: Checkout code
+        uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8 # ratchet:actions/checkout@v6
+        with:
+          persist-credentials: false
+
+      - name: Setup Python and Install Dependencies
+        uses: ./.github/actions/setup-python-and-install-dependencies
+        with:
+          requirements: |
+            backend/requirements/default.txt
+            backend/requirements/dev.txt
+
+      - name: Generate OpenAPI schema and Python client
+        shell: bash
+        run: |
+          ods openapi all
+
+      # needed for pulling external images otherwise, we hit the "Unauthenticated users" limit
+      # https://docs.docker.com/docker-hub/usage/
+      - name: Login to Docker Hub
+        uses: docker/login-action@5e57cd118135c172c3672efd75eb46360885c0ef # ratchet:docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKER_USERNAME }}
+          password: ${{ secrets.DOCKER_TOKEN }}
+
+      - name: Start Docker containers
+        working-directory: ./deployment/docker_compose
+        run: |
+          docker compose -f docker-compose.yml -f docker-compose.dev.yml up -d \
+            relational_db
+
+      - name: Run Database Tests
+        working-directory: ./backend
+        run: pytest -m alembic tests/integration/tests/migrations/

.github/workflows/pr-external-dependency-unit-tests.yml

@@ -38,6 +38,8 @@ env:
   # LLMs
   OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
   ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
+  VERTEX_CREDENTIALS: ${{ secrets.VERTEX_CREDENTIALS }}
+  VERTEX_LOCATION: ${{ vars.VERTEX_LOCATION }}
 
   # Code Interpreter
   # TODO: debug why this is failing and enable

.github/workflows/pr-helm-chart-testing.yml

@@ -6,11 +6,11 @@ concurrency:
 on:
   merge_group:
   pull_request:
-    branches: [ main ]
+    branches: [main]
   push:
     tags:
       - "v*.*.*"
-  workflow_dispatch:  # Allows manual triggering
+  workflow_dispatch: # Allows manual triggering
 
 permissions:
   contents: read
@@ -18,225 +18,233 @@ permissions:
 jobs:
   helm-chart-check:
     # See https://runs-on.com/runners/linux/
-    runs-on: [runs-on,runner=8cpu-linux-x64,hdd=256,"run-id=${{ github.run_id }}-helm-chart-check"]
+    runs-on:
+      [
+        runs-on,
+        runner=8cpu-linux-x64,
+        hdd=256,
+        "run-id=${{ github.run_id }}-helm-chart-check",
+      ]
     timeout-minutes: 45
 
     # fetch-depth 0 is required for helm/chart-testing-action
     steps:
-    - name: Checkout code
-      uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8 # ratchet:actions/checkout@v6
-      with:
-        fetch-depth: 0
-        persist-credentials: false
+      - name: Checkout code
+        uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8 # ratchet:actions/checkout@v6
+        with:
+          fetch-depth: 0
+          persist-credentials: false
 
-    - name: Set up Helm
-      uses: azure/setup-helm@1a275c3b69536ee54be43f2070a358922e12c8d4 # ratchet:azure/setup-helm@v4.3.1
-      with:
-        version: v3.19.0
+      - name: Set up Helm
+        uses: azure/setup-helm@1a275c3b69536ee54be43f2070a358922e12c8d4 # ratchet:azure/setup-helm@v4.3.1
+        with:
+          version: v3.19.0
 
-    - name: Set up chart-testing
-      uses: helm/chart-testing-action@6ec842c01de15ebb84c8627d2744a0c2f2755c9f # ratchet:helm/chart-testing-action@v2.8.0
+      - name: Set up chart-testing
+        # NOTE: This is Jamison's patch from https://github.com/helm/chart-testing-action/pull/194
+        uses: helm/chart-testing-action@8958a6ac472cbd8ee9a8fbb6f1acbc1b0e966e44 # zizmor: ignore[impostor-commit]
+        with:
+          uv_version: "0.9.9"
 
-    # even though we specify chart-dirs in ct.yaml, it isn't used by ct for the list-changed command...
-    - name: Run chart-testing (list-changed)
-      id: list-changed
-      env:
-        DEFAULT_BRANCH: ${{ github.event.repository.default_branch }}
-      run: |
-        echo "default_branch: ${DEFAULT_BRANCH}"
-        changed=$(ct list-changed --remote origin --target-branch ${DEFAULT_BRANCH} --chart-dirs deployment/helm/charts)
-        echo "list-changed output: $changed"
-        if [[ -n "$changed" ]]; then
-          echo "changed=true" >> "$GITHUB_OUTPUT"
-        fi
-
-    # uncomment to force run chart-testing
-#     - name: Force run chart-testing (list-changed)
-#       id: list-changed
-#       run: echo "changed=true" >> $GITHUB_OUTPUT
-
-    # lint all charts if any changes were detected
-    - name: Run chart-testing (lint)
-      if: steps.list-changed.outputs.changed == 'true'
-      run: ct lint --config ct.yaml --all
-      # the following would lint only changed charts, but linting isn't expensive
-      # run: ct lint --config ct.yaml --target-branch ${{ github.event.repository.default_branch }}
-
-    - name: Create kind cluster
-      if: steps.list-changed.outputs.changed == 'true'
-      uses: helm/kind-action@92086f6be054225fa813e0a4b13787fc9088faab # ratchet:helm/kind-action@v1.13.0
-
-    - name: Pre-install cluster status check
-      if: steps.list-changed.outputs.changed == 'true'
-      run: |
-        echo "=== Pre-install Cluster Status ==="
-        kubectl get nodes -o wide
-        kubectl get pods --all-namespaces
-        kubectl get storageclass
-
-    - name: Add Helm repositories and update
-      if: steps.list-changed.outputs.changed == 'true'
-      run: |
-        echo "=== Adding Helm repositories ==="
-        helm repo add ingress-nginx https://kubernetes.github.io/ingress-nginx
-        helm repo add vespa https://onyx-dot-app.github.io/vespa-helm-charts
-        helm repo add cloudnative-pg https://cloudnative-pg.github.io/charts
-        helm repo add ot-container-kit https://ot-container-kit.github.io/helm-charts
-        helm repo add minio https://charts.min.io/
-        helm repo add code-interpreter https://onyx-dot-app.github.io/code-interpreter/
-        helm repo update
-
-    - name: Install Redis operator
-      if: steps.list-changed.outputs.changed == 'true'
-      shell: bash
-      run: |
-        echo "=== Installing redis-operator CRDs ==="
-        helm upgrade --install redis-operator ot-container-kit/redis-operator \
-          --namespace redis-operator --create-namespace --wait --timeout 300s
-
-    - name: Pre-pull required images
-      if: steps.list-changed.outputs.changed == 'true'
-      run: |
-        echo "=== Pre-pulling required images to avoid timeout ==="
-        KIND_CLUSTER=$(kubectl config current-context | sed 's/kind-//')
-        echo "Kind cluster: $KIND_CLUSTER"
-
-        IMAGES=(
-          "ghcr.io/cloudnative-pg/cloudnative-pg:1.27.0"
-          "quay.io/opstree/redis:v7.0.15"
-          "docker.io/onyxdotapp/onyx-web-server:latest"
-        )
-
-        for image in "${IMAGES[@]}"; do
-          echo "Pre-pulling $image"
-          if docker pull "$image"; then
-            kind load docker-image "$image" --name "$KIND_CLUSTER" || echo "Failed to load $image into kind"
-          else
-            echo "Failed to pull $image"
+      # even though we specify chart-dirs in ct.yaml, it isn't used by ct for the list-changed command...
+      - name: Run chart-testing (list-changed)
+        id: list-changed
+        env:
+          DEFAULT_BRANCH: ${{ github.event.repository.default_branch }}
+        run: |
+          echo "default_branch: ${DEFAULT_BRANCH}"
+          changed=$(ct list-changed --remote origin --target-branch ${DEFAULT_BRANCH} --chart-dirs deployment/helm/charts)
+          echo "list-changed output: $changed"
+          if [[ -n "$changed" ]]; then
+            echo "changed=true" >> "$GITHUB_OUTPUT"
           fi
-        done
 
-        echo "=== Images loaded into Kind cluster ==="
-        docker exec "$KIND_CLUSTER"-control-plane crictl images | grep -E "(cloudnative-pg|redis|onyx)" || echo "Some images may still be loading..."
+      # uncomment to force run chart-testing
+      #     - name: Force run chart-testing (list-changed)
+      #       id: list-changed
+      #       run: echo "changed=true" >> $GITHUB_OUTPUT
+      # lint all charts if any changes were detected
+      - name: Run chart-testing (lint)
+        if: steps.list-changed.outputs.changed == 'true'
+        run: ct lint --config ct.yaml --all
+        # the following would lint only changed charts, but linting isn't expensive
+        # run: ct lint --config ct.yaml --target-branch ${{ github.event.repository.default_branch }}
 
-    - name: Validate chart dependencies
-      if: steps.list-changed.outputs.changed == 'true'
-      run: |
-        echo "=== Validating chart dependencies ==="
-        cd deployment/helm/charts/onyx
-        helm dependency update
-        helm lint .
+      - name: Create kind cluster
+        if: steps.list-changed.outputs.changed == 'true'
+        uses: helm/kind-action@92086f6be054225fa813e0a4b13787fc9088faab # ratchet:helm/kind-action@v1.13.0
 
-    - name: Run chart-testing (install) with enhanced monitoring
-      timeout-minutes: 25
-      if: steps.list-changed.outputs.changed == 'true'
-      run: |
-        echo "=== Starting chart installation with monitoring ==="
+      - name: Pre-install cluster status check
+        if: steps.list-changed.outputs.changed == 'true'
+        run: |
+          echo "=== Pre-install Cluster Status ==="
+          kubectl get nodes -o wide
+          kubectl get pods --all-namespaces
+          kubectl get storageclass
 
-        # Function to monitor cluster state
-        monitor_cluster() {
-          while true; do
-            echo "=== Cluster Status Check at $(date) ==="
-            # Only show non-running pods to reduce noise
-            NON_RUNNING_PODS=$(kubectl get pods --all-namespaces --field-selector=status.phase!=Running,status.phase!=Succeeded --no-headers 2>/dev/null | wc -l)
-            if [ "$NON_RUNNING_PODS" -gt 0 ]; then
-              echo "Non-running pods:"
-              kubectl get pods --all-namespaces --field-selector=status.phase!=Running,status.phase!=Succeeded
+      - name: Add Helm repositories and update
+        if: steps.list-changed.outputs.changed == 'true'
+        run: |
+          echo "=== Adding Helm repositories ==="
+          helm repo add ingress-nginx https://kubernetes.github.io/ingress-nginx
+          helm repo add vespa https://onyx-dot-app.github.io/vespa-helm-charts
+          helm repo add cloudnative-pg https://cloudnative-pg.github.io/charts
+          helm repo add ot-container-kit https://ot-container-kit.github.io/helm-charts
+          helm repo add minio https://charts.min.io/
+          helm repo add code-interpreter https://onyx-dot-app.github.io/code-interpreter/
+          helm repo update
+
+      - name: Install Redis operator
+        if: steps.list-changed.outputs.changed == 'true'
+        shell: bash
+        run: |
+          echo "=== Installing redis-operator CRDs ==="
+          helm upgrade --install redis-operator ot-container-kit/redis-operator \
+            --namespace redis-operator --create-namespace --wait --timeout 300s
+
+      - name: Pre-pull required images
+        if: steps.list-changed.outputs.changed == 'true'
+        run: |
+          echo "=== Pre-pulling required images to avoid timeout ==="
+          KIND_CLUSTER=$(kubectl config current-context | sed 's/kind-//')
+          echo "Kind cluster: $KIND_CLUSTER"
+
+          IMAGES=(
+            "ghcr.io/cloudnative-pg/cloudnative-pg:1.27.0"
+            "quay.io/opstree/redis:v7.0.15"
+            "docker.io/onyxdotapp/onyx-web-server:latest"
+          )
+
+          for image in "${IMAGES[@]}"; do
+            echo "Pre-pulling $image"
+            if docker pull "$image"; then
+              kind load docker-image "$image" --name "$KIND_CLUSTER" || echo "Failed to load $image into kind"
             else
-              echo "All pods running successfully"
+              echo "Failed to pull $image"
             fi
-            # Only show recent events if there are issues
-            RECENT_EVENTS=$(kubectl get events --sort-by=.lastTimestamp --all-namespaces --field-selector=type!=Normal 2>/dev/null | tail -5)
-            if [ -n "$RECENT_EVENTS" ]; then
-              echo "Recent warnings/errors:"
-              echo "$RECENT_EVENTS"
-            fi
-            sleep 60
           done
-        }
 
-        # Start monitoring in background
-        monitor_cluster &
-        MONITOR_PID=$!
+          echo "=== Images loaded into Kind cluster ==="
+          docker exec "$KIND_CLUSTER"-control-plane crictl images | grep -E "(cloudnative-pg|redis|onyx)" || echo "Some images may still be loading..."
 
-        # Set up cleanup
-        cleanup() {
-          echo "=== Cleaning up monitoring process ==="
-          kill $MONITOR_PID 2>/dev/null || true
+      - name: Validate chart dependencies
+        if: steps.list-changed.outputs.changed == 'true'
+        run: |
+          echo "=== Validating chart dependencies ==="
+          cd deployment/helm/charts/onyx
+          helm dependency update
+          helm lint .
+
+      - name: Run chart-testing (install) with enhanced monitoring
+        timeout-minutes: 25
+        if: steps.list-changed.outputs.changed == 'true'
+        run: |
+          echo "=== Starting chart installation with monitoring ==="
+
+          # Function to monitor cluster state
+          monitor_cluster() {
+            while true; do
+              echo "=== Cluster Status Check at $(date) ==="
+              # Only show non-running pods to reduce noise
+              NON_RUNNING_PODS=$(kubectl get pods --all-namespaces --field-selector=status.phase!=Running,status.phase!=Succeeded --no-headers 2>/dev/null | wc -l)
+              if [ "$NON_RUNNING_PODS" -gt 0 ]; then
+                echo "Non-running pods:"
+                kubectl get pods --all-namespaces --field-selector=status.phase!=Running,status.phase!=Succeeded
+              else
+                echo "All pods running successfully"
+              fi
+              # Only show recent events if there are issues
+              RECENT_EVENTS=$(kubectl get events --sort-by=.lastTimestamp --all-namespaces --field-selector=type!=Normal 2>/dev/null | tail -5)
+              if [ -n "$RECENT_EVENTS" ]; then
+                echo "Recent warnings/errors:"
+                echo "$RECENT_EVENTS"
+              fi
+              sleep 60
+            done
+          }
+
+          # Start monitoring in background
+          monitor_cluster &
+          MONITOR_PID=$!
+
+          # Set up cleanup
+          cleanup() {
+            echo "=== Cleaning up monitoring process ==="
+            kill $MONITOR_PID 2>/dev/null || true
+            echo "=== Final cluster state ==="
+            kubectl get pods --all-namespaces
+            kubectl get events --all-namespaces --sort-by=.lastTimestamp | tail -20
+          }
+
+          # Trap cleanup on exit
+          trap cleanup EXIT
+
+          # Run the actual installation with detailed logging
+          echo "=== Starting ct install ==="
+          set +e
+          ct install --all \
+            --helm-extra-set-args="\
+              --set=nginx.enabled=false \
+              --set=minio.enabled=false \
+              --set=vespa.enabled=false \
+              --set=slackbot.enabled=false \
+              --set=postgresql.enabled=true \
+              --set=postgresql.nameOverride=cloudnative-pg \
+              --set=postgresql.cluster.storage.storageClass=standard \
+              --set=redis.enabled=true \
+              --set=redis.storageSpec.volumeClaimTemplate.spec.storageClassName=standard \
+              --set=webserver.replicaCount=1 \
+              --set=api.replicaCount=0 \
+              --set=inferenceCapability.replicaCount=0 \
+              --set=indexCapability.replicaCount=0 \
+              --set=celery_beat.replicaCount=0 \
+              --set=celery_worker_heavy.replicaCount=0 \
+              --set=celery_worker_docfetching.replicaCount=0 \
+              --set=celery_worker_docprocessing.replicaCount=0 \
+              --set=celery_worker_light.replicaCount=0 \
+              --set=celery_worker_monitoring.replicaCount=0 \
+              --set=celery_worker_primary.replicaCount=0 \
+              --set=celery_worker_user_file_processing.replicaCount=0 \
+              --set=celery_worker_user_files_indexing.replicaCount=0" \
+            --helm-extra-args="--timeout 900s --debug" \
+            --debug --config ct.yaml
+          CT_EXIT=$?
+          set -e
+
+          if [[ $CT_EXIT -ne 0 ]]; then
+            echo "ct install failed with exit code $CT_EXIT"
+            exit $CT_EXIT
+          else
+            echo "=== Installation completed successfully ==="
+          fi
+
+          kubectl get pods --all-namespaces
+
+      - name: Post-install verification
+        if: steps.list-changed.outputs.changed == 'true'
+        run: |
+          echo "=== Post-install verification ==="
+          kubectl get pods --all-namespaces
+          kubectl get services --all-namespaces
+          # Only show issues if they exist
+          kubectl describe pods --all-namespaces | grep -A 5 -B 2 "Failed\|Error\|Warning" || echo "No pod issues found"
+
+      - name: Cleanup on failure
+        if: failure() && steps.list-changed.outputs.changed == 'true'
+        run: |
+          echo "=== Cleanup on failure ==="
           echo "=== Final cluster state ==="
           kubectl get pods --all-namespaces
-          kubectl get events --all-namespaces --sort-by=.lastTimestamp | tail -20
-        }
+          kubectl get events --all-namespaces --sort-by=.lastTimestamp | tail -10
 
-        # Trap cleanup on exit
-        trap cleanup EXIT
+          echo "=== Pod descriptions for debugging ==="
+          kubectl describe pods --all-namespaces | grep -A 10 -B 3 "Failed\|Error\|Warning\|Pending" || echo "No problematic pods found"
 
-        # Run the actual installation with detailed logging
-        echo "=== Starting ct install ==="
-        set +e
-        ct install --all \
-          --helm-extra-set-args="\
-            --set=nginx.enabled=false \
-            --set=minio.enabled=false \
-            --set=vespa.enabled=false \
-            --set=slackbot.enabled=false \
-            --set=postgresql.enabled=true \
-            --set=postgresql.nameOverride=cloudnative-pg \
-            --set=postgresql.cluster.storage.storageClass=standard \
-            --set=redis.enabled=true \
-            --set=redis.storageSpec.volumeClaimTemplate.spec.storageClassName=standard \
-            --set=webserver.replicaCount=1 \
-            --set=api.replicaCount=0 \
-            --set=inferenceCapability.replicaCount=0 \
-            --set=indexCapability.replicaCount=0 \
-            --set=celery_beat.replicaCount=0 \
-            --set=celery_worker_heavy.replicaCount=0 \
-            --set=celery_worker_docfetching.replicaCount=0 \
-            --set=celery_worker_docprocessing.replicaCount=0 \
-            --set=celery_worker_light.replicaCount=0 \
-            --set=celery_worker_monitoring.replicaCount=0 \
-            --set=celery_worker_primary.replicaCount=0 \
-            --set=celery_worker_user_file_processing.replicaCount=0 \
-            --set=celery_worker_user_files_indexing.replicaCount=0" \
-          --helm-extra-args="--timeout 900s --debug" \
-          --debug --config ct.yaml
-        CT_EXIT=$?
-        set -e
+          echo "=== Recent logs for debugging ==="
+          kubectl logs --all-namespaces --tail=50 | grep -i "error\|timeout\|failed\|pull" || echo "No error logs found"
 
-        if [[ $CT_EXIT -ne 0 ]]; then
-          echo "ct install failed with exit code $CT_EXIT"
-          exit $CT_EXIT
-        else
-          echo "=== Installation completed successfully ==="
-        fi
-
-        kubectl get pods --all-namespaces
-
-    - name: Post-install verification
-      if: steps.list-changed.outputs.changed == 'true'
-      run: |
-        echo "=== Post-install verification ==="
-        kubectl get pods --all-namespaces
-        kubectl get services --all-namespaces
-        # Only show issues if they exist
-        kubectl describe pods --all-namespaces | grep -A 5 -B 2 "Failed\|Error\|Warning" || echo "No pod issues found"
-
-    - name: Cleanup on failure
-      if: failure() && steps.list-changed.outputs.changed == 'true'
-      run: |
-        echo "=== Cleanup on failure ==="
-        echo "=== Final cluster state ==="
-        kubectl get pods --all-namespaces
-        kubectl get events --all-namespaces --sort-by=.lastTimestamp | tail -10
-
-        echo "=== Pod descriptions for debugging ==="
-        kubectl describe pods --all-namespaces | grep -A 10 -B 3 "Failed\|Error\|Warning\|Pending" || echo "No problematic pods found"
-
-        echo "=== Recent logs for debugging ==="
-        kubectl logs --all-namespaces --tail=50 | grep -i "error\|timeout\|failed\|pull" || echo "No error logs found"
-
-        echo "=== Helm releases ==="
-        helm list --all-namespaces
-      # the following would install only changed charts, but we only have one chart so
-      # don't worry about that for now
-      # run: ct install --target-branch ${{ github.event.repository.default_branch }}
+          echo "=== Helm releases ==="
+          helm list --all-namespaces
+        # the following would install only changed charts, but we only have one chart so
+        # don't worry about that for now
+        # run: ct install --target-branch ${{ github.event.repository.default_branch }}

.github/workflows/pr-integration-tests.yml

@@ -56,7 +56,7 @@ jobs:
         id: set-matrix
         run: |
           # Find all leaf-level directories in both test directories
-          tests_dirs=$(find backend/tests/integration/tests -mindepth 1 -maxdepth 1 -type d ! -name "__pycache__" -exec basename {} \; | sort)
+          tests_dirs=$(find backend/tests/integration/tests -mindepth 1 -maxdepth 1 -type d ! -name "__pycache__" ! -name "mcp" -exec basename {} \; | sort)
           connector_dirs=$(find backend/tests/integration/connector_job_tests -mindepth 1 -maxdepth 1 -type d ! -name "__pycache__" -exec basename {} \; | sort)
 
           # Create JSON array with directory info
@@ -72,9 +72,14 @@ jobs:
           all_dirs="[${all_dirs%,}]"
           echo "test-dirs=$all_dirs" >> $GITHUB_OUTPUT
 
-
   build-backend-image:
-    runs-on: [runs-on, runner=1cpu-linux-arm64, "run-id=${{ github.run_id }}-build-backend-image", "extras=ecr-cache"]
+    runs-on:
+      [
+        runs-on,
+        runner=1cpu-linux-arm64,
+        "run-id=${{ github.run_id }}-build-backend-image",
+        "extras=ecr-cache",
+      ]
     timeout-minutes: 45
     steps:
       - uses: runs-on/action@cd2b598b0515d39d78c38a02d529db87d2196d1e # ratchet:runs-on/action@v2
@@ -127,9 +132,14 @@ jobs:
             type=registry,ref=${{ env.RUNS_ON_ECR_CACHE }}:backend-cache,mode=max
           no-cache: ${{ vars.DOCKER_NO_CACHE == 'true' }}
 
-
   build-model-server-image:
-    runs-on: [runs-on, runner=1cpu-linux-arm64, "run-id=${{ github.run_id }}-build-model-server-image", "extras=ecr-cache"]
+    runs-on:
+      [
+        runs-on,
+        runner=1cpu-linux-arm64,
+        "run-id=${{ github.run_id }}-build-model-server-image",
+        "extras=ecr-cache",
+      ]
     timeout-minutes: 45
     steps:
       - uses: runs-on/action@cd2b598b0515d39d78c38a02d529db87d2196d1e # ratchet:runs-on/action@v2
@@ -181,9 +191,14 @@ jobs:
             type=registry,ref=${{ env.RUNS_ON_ECR_CACHE }}:model-server-cache-${{ steps.format-branch.outputs.cache-suffix }},mode=max
             type=registry,ref=${{ env.RUNS_ON_ECR_CACHE }}:model-server-cache,mode=max
 
-
   build-integration-image:
-    runs-on: [runs-on, runner=2cpu-linux-arm64, "run-id=${{ github.run_id }}-build-integration-image", "extras=ecr-cache"]
+    runs-on:
+      [
+        runs-on,
+        runner=2cpu-linux-arm64,
+        "run-id=${{ github.run_id }}-build-integration-image",
+        "extras=ecr-cache",
+      ]
     timeout-minutes: 45
     steps:
       - uses: runs-on/action@cd2b598b0515d39d78c38a02d529db87d2196d1e # ratchet:runs-on/action@v2
@@ -225,7 +240,7 @@ jobs:
           CACHE_SUFFIX: ${{ steps.format-branch.outputs.cache-suffix }}
           HEAD_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
         run: |
-          cd backend && docker buildx bake --push \
+          docker buildx bake --push \
             --set backend.cache-from=type=registry,ref=${RUNS_ON_ECR_CACHE}:backend-cache-${HEAD_SHA} \
             --set backend.cache-from=type=registry,ref=${RUNS_ON_ECR_CACHE}:backend-cache-${CACHE_SUFFIX} \
             --set backend.cache-from=type=registry,ref=${RUNS_ON_ECR_CACHE}:backend-cache \
@@ -295,6 +310,7 @@ jobs:
           ONYX_MODEL_SERVER_IMAGE=${ECR_CACHE}:integration-test-model-server-test-${RUN_ID}
           INTEGRATION_TESTS_MODE=true
           CHECK_TTL_MANAGEMENT_TASK_FREQUENCY_IN_HOURS=0.001
+          AUTO_LLM_UPDATE_INTERVAL_SECONDS=1
           MCP_SERVER_ENABLED=true
           EOF
 
@@ -309,7 +325,6 @@ jobs:
             api_server \
             inference_model_server \
             indexing_model_server \
-            mcp_server \
             background \
             -d
         id: start_docker
@@ -352,12 +367,6 @@ jobs:
           }
 
           wait_for_service "http://localhost:8080/health" "API server"
-          test_dir="${{ matrix.test-dir.path }}"
-          if [ "$test_dir" = "tests/mcp" ]; then
-            wait_for_service "http://localhost:8090/health" "MCP server"
-          else
-            echo "Skipping MCP server wait for non-MCP suite: $test_dir"
-          fi
           echo "Finished waiting for services."
 
       - name: Start Mock Services
@@ -387,8 +396,6 @@ jobs:
               -e VESPA_HOST=index \
               -e REDIS_HOST=cache \
               -e API_SERVER_HOST=api_server \
-              -e MCP_SERVER_HOST=mcp_server \
-              -e MCP_SERVER_PORT=8090 \
               -e OPENAI_API_KEY=${OPENAI_API_KEY} \
               -e EXA_API_KEY=${EXA_API_KEY} \
               -e SLACK_BOT_TOKEN=${SLACK_BOT_TOKEN} \
@@ -437,15 +444,16 @@ jobs:
           path: ${{ github.workspace }}/docker-compose.log
       # ------------------------------------------------------------
 
-
   multitenant-tests:
     needs:
+      [build-backend-image, build-model-server-image, build-integration-image]
+    runs-on:
       [
-        build-backend-image,
-        build-model-server-image,
-        build-integration-image,
+        runs-on,
+        runner=8cpu-linux-arm64,
+        "run-id=${{ github.run_id }}-multitenant-tests",
+        "extras=ecr-cache",
       ]
-    runs-on: [runs-on, runner=8cpu-linux-arm64, "run-id=${{ github.run_id }}-multitenant-tests", "extras=ecr-cache"]
     timeout-minutes: 45
 
     steps:
@@ -472,10 +480,10 @@ jobs:
           AUTH_TYPE=cloud \
           REQUIRE_EMAIL_VERIFICATION=false \
           DISABLE_TELEMETRY=true \
+          OPENAI_DEFAULT_API_KEY=${OPENAI_API_KEY} \
           ONYX_BACKEND_IMAGE=${ECR_CACHE}:integration-test-backend-test-${RUN_ID} \
           ONYX_MODEL_SERVER_IMAGE=${ECR_CACHE}:integration-test-model-server-test-${RUN_ID} \
           DEV_MODE=true \
-          MCP_SERVER_ENABLED=true \
           docker compose -f docker-compose.multitenant-dev.yml up \
             relational_db \
             index \
@@ -484,7 +492,6 @@ jobs:
             api_server \
             inference_model_server \
             indexing_model_server \
-            mcp_server \
             background \
             -d
         id: start_docker_multi_tenant
@@ -533,8 +540,6 @@ jobs:
             -e VESPA_HOST=index \
             -e REDIS_HOST=cache \
             -e API_SERVER_HOST=api_server \
-            -e MCP_SERVER_HOST=mcp_server \
-            -e MCP_SERVER_PORT=8090 \
             -e OPENAI_API_KEY=${OPENAI_API_KEY} \
             -e EXA_API_KEY=${EXA_API_KEY} \
             -e SLACK_BOT_TOKEN=${SLACK_BOT_TOKEN} \

.github/workflows/pr-mit-integration-tests.yml

@@ -48,7 +48,7 @@ jobs:
         id: set-matrix
         run: |
           # Find all leaf-level directories in both test directories
-          tests_dirs=$(find backend/tests/integration/tests -mindepth 1 -maxdepth 1 -type d ! -name "__pycache__" -exec basename {} \; | sort)
+          tests_dirs=$(find backend/tests/integration/tests -mindepth 1 -maxdepth 1 -type d ! -name "__pycache__" ! -name "mcp" -exec basename {} \; | sort)
           connector_dirs=$(find backend/tests/integration/connector_job_tests -mindepth 1 -maxdepth 1 -type d ! -name "__pycache__" -exec basename {} \; | sort)
 
           # Create JSON array with directory info
@@ -65,7 +65,13 @@ jobs:
           echo "test-dirs=$all_dirs" >> $GITHUB_OUTPUT
 
   build-backend-image:
-    runs-on: [runs-on, runner=1cpu-linux-arm64, "run-id=${{ github.run_id }}-build-backend-image", "extras=ecr-cache"]
+    runs-on:
+      [
+        runs-on,
+        runner=1cpu-linux-arm64,
+        "run-id=${{ github.run_id }}-build-backend-image",
+        "extras=ecr-cache",
+      ]
     timeout-minutes: 45
     steps:
       - uses: runs-on/action@cd2b598b0515d39d78c38a02d529db87d2196d1e # ratchet:runs-on/action@v2
@@ -119,7 +125,13 @@ jobs:
           no-cache: ${{ vars.DOCKER_NO_CACHE == 'true' }}
 
   build-model-server-image:
-    runs-on: [runs-on, runner=1cpu-linux-arm64, "run-id=${{ github.run_id }}-build-model-server-image", "extras=ecr-cache"]
+    runs-on:
+      [
+        runs-on,
+        runner=1cpu-linux-arm64,
+        "run-id=${{ github.run_id }}-build-model-server-image",
+        "extras=ecr-cache",
+      ]
     timeout-minutes: 45
     steps:
       - uses: runs-on/action@cd2b598b0515d39d78c38a02d529db87d2196d1e # ratchet:runs-on/action@v2
@@ -172,7 +184,13 @@ jobs:
             type=registry,ref=${{ env.RUNS_ON_ECR_CACHE }}:model-server-cache,mode=max
 
   build-integration-image:
-    runs-on: [runs-on, runner=2cpu-linux-arm64, "run-id=${{ github.run_id }}-build-integration-image", "extras=ecr-cache"]
+    runs-on:
+      [
+        runs-on,
+        runner=2cpu-linux-arm64,
+        "run-id=${{ github.run_id }}-build-integration-image",
+        "extras=ecr-cache",
+      ]
     timeout-minutes: 45
     steps:
       - uses: runs-on/action@cd2b598b0515d39d78c38a02d529db87d2196d1e # ratchet:runs-on/action@v2
@@ -214,7 +232,7 @@ jobs:
           CACHE_SUFFIX: ${{ steps.format-branch.outputs.cache-suffix }}
           HEAD_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
         run: |
-          cd backend && docker buildx bake --push \
+          docker buildx bake --push \
             --set backend.cache-from=type=registry,ref=${RUNS_ON_ECR_CACHE}:backend-cache-${HEAD_SHA} \
             --set backend.cache-from=type=registry,ref=${RUNS_ON_ECR_CACHE}:backend-cache-${CACHE_SUFFIX} \
             --set backend.cache-from=type=registry,ref=${RUNS_ON_ECR_CACHE}:backend-cache \
@@ -283,6 +301,7 @@ jobs:
           ONYX_MODEL_SERVER_IMAGE=${ECR_CACHE}:integration-test-model-server-test-${RUN_ID}
           INTEGRATION_TESTS_MODE=true
           MCP_SERVER_ENABLED=true
+          AUTO_LLM_UPDATE_INTERVAL_SECONDS=1
           EOF
 
       - name: Start Docker containers
@@ -296,7 +315,6 @@ jobs:
             api_server \
             inference_model_server \
             indexing_model_server \
-            mcp_server \
             background \
             -d
         id: start_docker
@@ -339,12 +357,6 @@ jobs:
           }
 
           wait_for_service "http://localhost:8080/health" "API server"
-          test_dir="${{ matrix.test-dir.path }}"
-          if [ "$test_dir" = "tests/mcp" ]; then
-            wait_for_service "http://localhost:8090/health" "MCP server"
-          else
-            echo "Skipping MCP server wait for non-MCP suite: $test_dir"
-          fi
           echo "Finished waiting for services."
 
       - name: Start Mock Services
@@ -375,8 +387,6 @@ jobs:
               -e VESPA_HOST=index \
               -e REDIS_HOST=cache \
               -e API_SERVER_HOST=api_server \
-              -e MCP_SERVER_HOST=mcp_server \
-              -e MCP_SERVER_PORT=8090 \
               -e OPENAI_API_KEY=${OPENAI_API_KEY} \
               -e EXA_API_KEY=${EXA_API_KEY} \
               -e SLACK_BOT_TOKEN=${SLACK_BOT_TOKEN} \
@@ -420,7 +430,6 @@ jobs:
           path: ${{ github.workspace }}/docker-compose.log
       # ------------------------------------------------------------
 
-
   required:
     # NOTE: Github-hosted runners have about 20s faster queue times and are preferred here.
     runs-on: ubuntu-slim

.github/workflows/pr-playwright-tests.yml

@@ -54,7 +54,13 @@ env:
 
 jobs:
   build-web-image:
-    runs-on: [runs-on, runner=4cpu-linux-arm64, "run-id=${{ github.run_id }}-build-web-image", "extras=ecr-cache"]
+    runs-on:
+      [
+        runs-on,
+        runner=4cpu-linux-arm64,
+        "run-id=${{ github.run_id }}-build-web-image",
+        "extras=ecr-cache",
+      ]
     timeout-minutes: 45
     steps:
       - uses: runs-on/action@cd2b598b0515d39d78c38a02d529db87d2196d1e # ratchet:runs-on/action@v2
@@ -109,7 +115,13 @@ jobs:
           no-cache: ${{ vars.DOCKER_NO_CACHE == 'true' }}
 
   build-backend-image:
-    runs-on: [runs-on, runner=1cpu-linux-arm64, "run-id=${{ github.run_id }}-build-backend-image", "extras=ecr-cache"]
+    runs-on:
+      [
+        runs-on,
+        runner=1cpu-linux-arm64,
+        "run-id=${{ github.run_id }}-build-backend-image",
+        "extras=ecr-cache",
+      ]
     timeout-minutes: 45
     steps:
       - uses: runs-on/action@cd2b598b0515d39d78c38a02d529db87d2196d1e # ratchet:runs-on/action@v2
@@ -164,7 +176,13 @@ jobs:
           no-cache: ${{ vars.DOCKER_NO_CACHE == 'true' }}
 
   build-model-server-image:
-    runs-on: [runs-on, runner=1cpu-linux-arm64, "run-id=${{ github.run_id }}-build-model-server-image", "extras=ecr-cache"]
+    runs-on:
+      [
+        runs-on,
+        runner=1cpu-linux-arm64,
+        "run-id=${{ github.run_id }}-build-model-server-image",
+        "extras=ecr-cache",
+      ]
     timeout-minutes: 45
     steps:
       - uses: runs-on/action@cd2b598b0515d39d78c38a02d529db87d2196d1e # ratchet:runs-on/action@v2
@@ -238,14 +256,13 @@ jobs:
       - name: Checkout code
         uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8 # ratchet:actions/checkout@v6
         with:
-          fetch-depth: 0
           persist-credentials: false
 
       - name: Setup node
         uses: actions/setup-node@395ad3262231945c25e8478fd5baf05154b1d79f # ratchet:actions/setup-node@v4
         with:
           node-version: 22
-          cache: 'npm'
+          cache: "npm"
           cache-dependency-path: ./web/package-lock.json
 
       - name: Install node dependencies
@@ -454,7 +471,6 @@ jobs:
         if: ${{ contains(needs.*.result, 'failure') || contains(needs.*.result, 'cancelled') || contains(needs.*.result, 'skipped') }}
         run: exit 1
 
-
 # NOTE: Chromatic UI diff testing is currently disabled.
 # We are using Playwright for local and CI testing without visual regression checks.
 # Chromatic may be reintroduced in the future for UI diff testing if needed.

.github/workflows/release-devtools.yml

@@ -16,21 +16,22 @@ jobs:
     strategy:
       matrix:
         os-arch:
-          - {goos: "linux", goarch: "amd64"}
-          - {goos: "linux", goarch: "arm64"}
-          - {goos: "windows", goarch: "amd64"}
-          - {goos: "windows", goarch: "arm64"}
-          - {goos: "darwin", goarch: "amd64"}
-          - {goos: "darwin", goarch: "arm64"}
-          - {goos: "", goarch: ""}
+          - { goos: "linux", goarch: "amd64" }
+          - { goos: "linux", goarch: "arm64" }
+          - { goos: "windows", goarch: "amd64" }
+          - { goos: "windows", goarch: "arm64" }
+          - { goos: "darwin", goarch: "amd64" }
+          - { goos: "darwin", goarch: "arm64" }
+          - { goos: "", goarch: "" }
     steps:
       - uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8 # ratchet:actions/checkout@v6
         with:
           persist-credentials: false
           fetch-depth: 0
-      - uses: astral-sh/setup-uv@1e862dfacbd1d6d858c55d9b792c756523627244 # ratchet:astral-sh/setup-uv@v7
+      - uses: astral-sh/setup-uv@ed21f2f24f8dd64503750218de024bcf64c7250a # ratchet:astral-sh/setup-uv@v7
         with:
           enable-cache: false
+          version: "0.9.9"
       - run: |
           GOOS="${{ matrix.os-arch.goos }}" \
           GOARCH="${{ matrix.os-arch.goarch }}" \

.github/workflows/zizmor.yml

@@ -21,17 +21,29 @@ jobs:
         with:
           persist-credentials: false
 
+      - name: Detect changes
+        id: filter
+        uses: dorny/paths-filter@de90cc6fb38fc0963ad72b210f1f284cd68cea36 # ratchet:dorny/paths-filter@v3
+        with:
+          filters: |
+            zizmor:
+              - '.github/**'
+
       - name: Install the latest version of uv
-        uses: astral-sh/setup-uv@1e862dfacbd1d6d858c55d9b792c756523627244 # ratchet:astral-sh/setup-uv@v7.1.4
+        if: steps.filter.outputs.zizmor == 'true' || github.ref_name == 'main'
+        uses: astral-sh/setup-uv@ed21f2f24f8dd64503750218de024bcf64c7250a # ratchet:astral-sh/setup-uv@v7
         with:
           enable-cache: false
+          version: "0.9.9"
 
       - name: Run zizmor
+        if: steps.filter.outputs.zizmor == 'true' || github.ref_name == 'main'
         run: uv run --no-sync --with zizmor zizmor --format=sarif . > results.sarif
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Upload SARIF file
+        if: steps.filter.outputs.zizmor == 'true' || github.ref_name == 'main'
         uses: github/codeql-action/upload-sarif@ba454b8ab46733eb6145342877cd148270bb77ab # ratchet:github/codeql-action/upload-sarif@codeql-bundle-v2.23.5
         with:
           sarif_file: results.sarif

.pre-commit-config.yaml

@@ -9,7 +9,7 @@ repos:
     rev: d30b4298e4fb63ce8609e29acdbcf4c9018a483c
     hooks:
       - id: uv-sync
-        args: ["--active", "--locked", "--all-extras"]
+        args: ["--locked", "--all-extras"]
       - id: uv-lock
         files: ^pyproject\.toml$
       - id: uv-export
@@ -75,12 +75,6 @@ repos:
       #   pass_filenames: true
       #   files: ^backend/.*\.py$
 
-  - repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: 3e8a8703264a2f4a69428a0aa4dcb512790b2c8c # frozen: v6.0.0
-    hooks:
-      - id: check-yaml
-        files: ^.github/
-
   - repo: https://github.com/rhysd/actionlint
     rev: a443f344ff32813837fa49f7aa6cbc478d770e62 # frozen: v1.7.9
     hooks:

.vscode/env_template.txt

@@ -1,36 +1,45 @@
-# Copy this file to .env in the .vscode folder
-# Fill in the <REPLACE THIS> values as needed, it is recommended to set the GEN_AI_API_KEY value to avoid having to set up an LLM in the UI
-# Also check out onyx/backend/scripts/restart_containers.sh for a script to restart the containers which Onyx relies on outside of VSCode/Cursor processes
+# Copy this file to .env in the .vscode folder.
+# Fill in the <REPLACE THIS> values as needed; it is recommended to set the
+# GEN_AI_API_KEY value to avoid having to set up an LLM in the UI.
+# Also check out onyx/backend/scripts/restart_containers.sh for a script to
+# restart the containers which Onyx relies on outside of VSCode/Cursor
+# processes.
 
-# For local dev, often user Authentication is not needed
+
+# For local dev, often user Authentication is not needed.
 AUTH_TYPE=disabled
 
-# Always keep these on for Dev
-# Logs model prompts, reasoning, and answer to stdout
+
+# Always keep these on for Dev.
+# Logs model prompts, reasoning, and answer to stdout.
 LOG_ONYX_MODEL_INTERACTIONS=True
 # More verbose logging
 LOG_LEVEL=debug
 
 
-# This passes top N results to LLM an additional time for reranking prior to answer generation
-# This step is quite heavy on token usage so we disable it for dev generally
+# This passes top N results to LLM an additional time for reranking prior to
+# answer generation.
+# This step is quite heavy on token usage so we disable it for dev generally.
 DISABLE_LLM_DOC_RELEVANCE=False
 
 
-# Useful if you want to toggle auth on/off (google_oauth/OIDC specifically)
+# Useful if you want to toggle auth on/off (google_oauth/OIDC specifically).
 OAUTH_CLIENT_ID=<REPLACE THIS>
 OAUTH_CLIENT_SECRET=<REPLACE THIS>
 OPENID_CONFIG_URL=<REPLACE THIS>
 SAML_CONF_DIR=/<ABSOLUTE PATH TO ONYX>/onyx/backend/ee/onyx/configs/saml_config
 
-# Generally not useful for dev, we don't generally want to set up an SMTP server for dev
+
+# Generally not useful for dev, we don't generally want to set up an SMTP server
+# for dev.
 REQUIRE_EMAIL_VERIFICATION=False
 
 
-# Set these so if you wipe the DB, you don't end up having to go through the UI every time
+# Set these so if you wipe the DB, you don't end up having to go through the UI
+# every time.
 GEN_AI_API_KEY=<REPLACE THIS>
 OPENAI_API_KEY=<REPLACE THIS>
-# If answer quality isn't important for dev, use gpt-4o-mini since it's cheaper
+# If answer quality isn't important for dev, use gpt-4o-mini since it's cheaper.
 GEN_AI_MODEL_VERSION=gpt-4o
 FAST_GEN_AI_MODEL_VERSION=gpt-4o
 
@@ -40,26 +49,36 @@ PYTHONPATH=../backend
 PYTHONUNBUFFERED=1
 
 
-# Enable the full set of Danswer Enterprise Edition features
-# NOTE: DO NOT ENABLE THIS UNLESS YOU HAVE A PAID ENTERPRISE LICENSE (or if you are using this for local testing/development)
+# Enable the full set of Danswer Enterprise Edition features.
+# NOTE: DO NOT ENABLE THIS UNLESS YOU HAVE A PAID ENTERPRISE LICENSE (or if you
+# are using this for local testing/development).
 ENABLE_PAID_ENTERPRISE_EDITION_FEATURES=False
 
+
 # S3 File Store Configuration (MinIO for local development)
 S3_ENDPOINT_URL=http://localhost:9004
 S3_FILE_STORE_BUCKET_NAME=onyx-file-store-bucket
 S3_AWS_ACCESS_KEY_ID=minioadmin
 S3_AWS_SECRET_ACCESS_KEY=minioadmin
 
-# Show extra/uncommon connectors
+
+# Show extra/uncommon connectors.
 SHOW_EXTRA_CONNECTORS=True
 
+
 # Local langsmith tracing
 LANGSMITH_TRACING="true"
 LANGSMITH_ENDPOINT="https://api.smith.langchain.com"
 LANGSMITH_API_KEY=<REPLACE_THIS>
 LANGSMITH_PROJECT=<REPLACE_THIS>
 
+
 # Local Confluence OAuth testing
 # OAUTH_CONFLUENCE_CLOUD_CLIENT_ID=<REPLACE_THIS>
 # OAUTH_CONFLUENCE_CLOUD_CLIENT_SECRET=<REPLACE_THIS>
-# NEXT_PUBLIC_TEST_ENV=True
+# NEXT_PUBLIC_TEST_ENV=True
+
+
+# OpenSearch
+# Arbitrary password is fine for local development.
+OPENSEARCH_INITIAL_ADMIN_PASSWORD=<REPLACE THIS>

.vscode/launch.template.jsonc

@@ -512,6 +512,21 @@
         "group": "3"
       }
     },
+    {
+      "name": "Clear and Restart OpenSearch Container",
+      // Generic debugger type, required arg but has no bearing on bash.
+      "type": "node",
+      "request": "launch",
+      "runtimeExecutable": "bash",
+      "runtimeArgs": [
+        "${workspaceFolder}/backend/scripts/restart_opensearch_container.sh"
+      ],
+      "cwd": "${workspaceFolder}",
+      "console": "integratedTerminal",
+      "presentation": {
+        "group": "3"
+      }
+    },
     {
       "name": "Eval CLI",
       "type": "debugpy",

AGENTS.md.template

@@ -1,13 +1,13 @@
 # AGENTS.md
 
-This file provides guidance to Codex when working with code in this repository.
+This file provides guidance to AI agents when working with code in this repository.
 
 ## KEY NOTES
 
-- If you run into any missing python dependency errors, try running your command with `source backend/.venv/bin/activate` \
+- If you run into any missing python dependency errors, try running your command with `source .venv/bin/activate` \
 to assume the python venv.
 - To make tests work, check the `.env` file at the root of the project to find an OpenAI key.
-- If using `playwright` to explore the frontend, you can usually log in with username `a@test.com` and password
+- If using `playwright` to explore the frontend, you can usually log in with username `a@example.com` and password
 `a`. The app can be accessed at `http://localhost:3000`.
 - You should assume that all Onyx services are running. To verify, you can check the `backend/log` directory to
 make sure we see logs coming out from the relevant service.
@@ -181,6 +181,286 @@ web/
 └── src/lib/                     # Utilities & business logic
 ```
 
+## Frontend Standards
+
+### 1. Import Standards
+
+**Always use absolute imports with the `@` prefix.**
+
+**Reason:** Moving files around becomes easier since you don't also have to update those import statements. This makes modifications to the codebase much nicer.
+
+```typescript
+// ✅ Good
+import { Button } from "@/components/ui/button";
+import { useAuth } from "@/hooks/useAuth";
+import { Text } from "@/refresh-components/texts/Text";
+
+// ❌ Bad
+import { Button } from "../../../components/ui/button";
+import { useAuth } from "./hooks/useAuth";
+```
+
+### 2. React Component Functions
+
+**Prefer regular functions over arrow functions for React components.**
+
+**Reason:** Functions just become easier to read.
+
+```typescript
+// ✅ Good
+function UserProfile({ userId }: UserProfileProps) {
+  return <div>User Profile</div>
+}
+
+// ❌ Bad
+const UserProfile = ({ userId }: UserProfileProps) => {
+  return <div>User Profile</div>
+}
+```
+
+### 3. Props Interface Extraction
+
+**Extract prop types into their own interface definitions.**
+
+**Reason:** Functions just become easier to read.
+
+```typescript
+// ✅ Good
+interface UserCardProps {
+  user: User
+  showActions?: boolean
+  onEdit?: (userId: string) => void
+}
+
+function UserCard({ user, showActions = false, onEdit }: UserCardProps) {
+  return <div>User Card</div>
+}
+
+// ❌ Bad
+function UserCard({
+  user,
+  showActions = false,
+  onEdit
+}: {
+  user: User
+  showActions?: boolean
+  onEdit?: (userId: string) => void
+}) {
+  return <div>User Card</div>
+}
+```
+
+### 4. Spacing Guidelines
+
+**Prefer padding over margins for spacing.**
+
+**Reason:** We want to consolidate usage to paddings instead of margins.
+
+```typescript
+// ✅ Good
+<div className="p-4 space-y-2">
+  <div className="p-2">Content</div>
+</div>
+
+// ❌ Bad
+<div className="m-4 space-y-2">
+  <div className="m-2">Content</div>
+</div>
+```
+
+### 5. Tailwind Dark Mode
+
+**Strictly forbid using the `dark:` modifier in Tailwind classes, except for logo icon handling.**
+
+**Reason:** The `colors.css` file already, VERY CAREFULLY, defines what the exact opposite colour of each light-mode colour is. Overriding this behaviour is VERY bad and will lead to horrible UI breakages.
+
+**Exception:** The `createLogoIcon` helper in `web/src/components/icons/icons.tsx` uses `dark:` modifiers (`dark:invert`, `dark:hidden`, `dark:block`) to handle third-party logo icons that cannot automatically adapt through `colors.css`. This is the ONLY acceptable use of dark mode modifiers.
+
+```typescript
+// ✅ Good - Standard components use `web/tailwind-themes/tailwind.config.js` / `web/src/app/css/colors.css`
+<div className="bg-background-neutral-03 text-text-02">
+  Content
+</div>
+
+// ✅ Good - Logo icons with dark mode handling via createLogoIcon
+export const GithubIcon = createLogoIcon(githubLightIcon, {
+  monochromatic: true,  // Will apply dark:invert internally
+});
+
+export const GitbookIcon = createLogoIcon(gitbookLightIcon, {
+  darkSrc: gitbookDarkIcon,  // Will use dark:hidden/dark:block internally
+});
+
+// ❌ Bad - Manual dark mode overrides
+<div className="bg-white dark:bg-black text-black dark:text-white">
+  Content
+</div>
+```
+
+### 6. Class Name Utilities
+
+**Use the `cn` utility instead of raw string formatting for classNames.**
+
+**Reason:** `cn`s are easier to read. They also allow for more complex types (i.e., string-arrays) to get formatted properly (it flattens each element in that string array down). As a result, it can allow things such as conditionals (i.e., `myCondition && "some-tailwind-class"`, which evaluates to `false` when `myCondition` is `false`) to get filtered out.
+
+```typescript
+import { cn } from '@/lib/utils'
+
+// ✅ Good
+<div className={cn(
+  'base-class',
+  isActive && 'active-class',
+  className
+)}>
+  Content
+</div>
+
+// ❌ Bad
+<div className={`base-class ${isActive ? 'active-class' : ''} ${className}`}>
+  Content
+</div>
+```
+
+### 7. Custom Hooks Organization
+
+**Follow a "hook-per-file" layout. Each hook should live in its own file within `web/src/hooks`.**
+
+**Reason:** This is just a layout preference. Keeps code clean.
+
+```typescript
+// web/src/hooks/useUserData.ts
+export function useUserData(userId: string) {
+  // hook implementation
+}
+
+// web/src/hooks/useLocalStorage.ts
+export function useLocalStorage<T>(key: string, initialValue: T) {
+  // hook implementation
+}
+```
+
+### 8. Icon Usage
+
+**ONLY use icons from the `web/src/icons` directory. Do NOT use icons from `react-icons`, `lucide`, or other external libraries.**
+
+**Reason:** We have a very carefully curated selection of icons that match our Onyx guidelines. We do NOT want to muddy those up with different aesthetic stylings.
+
+```typescript
+// ✅ Good
+import SvgX from "@/icons/x";
+import SvgMoreHorizontal from "@/icons/more-horizontal";
+
+// ❌ Bad
+import { User } from "lucide-react";
+import { FiSearch } from "react-icons/fi";
+```
+
+**Missing Icons**: If an icon is needed but doesn't exist in the `web/src/icons` directory, import it from Figma using the Figma MCP tool and add it to the icons directory.
+If you need help with this step, reach out to `raunak@onyx.app`.
+
+### 9. Text Rendering
+
+**Prefer using the `refresh-components/texts/Text` component for all text rendering. Avoid "naked" text nodes.**
+
+**Reason:** The `Text` component is fully compliant with the stylings provided in Figma. It provides easy utilities to specify the text-colour and font-size in the form of flags. Super duper easy.
+
+```typescript
+// ✅ Good
+import { Text } from '@/refresh-components/texts/Text'
+
+function UserCard({ name }: { name: string }) {
+  return (
+    <Text
+      {/* The `text03` flag makes the text it renders to be coloured the 3rd-scale grey */}
+      text03
+      {/* The `mainAction` flag makes the text it renders to be "main-action" font + line-height + weightage, as described in the Figma */}
+      mainAction
+    >
+      {name}
+    </Text>
+  )
+}
+
+// ❌ Bad
+function UserCard({ name }: { name: string }) {
+  return (
+    <div>
+      <h2>{name}</h2>
+      <p>User details</p>
+    </div>
+  )
+}
+```
+
+### 10. Component Usage
+
+**Heavily avoid raw HTML input components. Always use components from the `web/src/refresh-components` or `web/lib/opal/src` directory.**
+
+**Reason:** We've put in a lot of effort to unify the components that are rendered in the Onyx app. Using raw components breaks the entire UI of the application, and leaves it in a muddier state than before.
+
+```typescript
+// ✅ Good
+import Button from '@/refresh-components/buttons/Button'
+import InputTypeIn from '@/refresh-components/inputs/InputTypeIn'
+import SvgPlusCircle from '@/icons/plus-circle'
+
+function ContactForm() {
+  return (
+    <form>
+      <InputTypeIn placeholder="Search..." />
+      <Button type="submit" leftIcon={SvgPlusCircle}>Submit</Button>
+    </form>
+  )
+}
+
+// ❌ Bad
+function ContactForm() {
+  return (
+    <form>
+      <input placeholder="Name" />
+      <textarea placeholder="Message" />
+      <button type="submit">Submit</button>
+    </form>
+  )
+}
+```
+
+### 11. Colors
+
+**Always use custom overrides for colors and borders rather than built in Tailwind CSS colors. These overrides live in `web/tailwind-themes/tailwind.config.js`.**
+
+**Reason:** Our custom color system uses CSS variables that automatically handle dark mode and maintain design consistency across the app. Standard Tailwind colors bypass this system.
+
+**Available color categories:**
+- **Text:** `text-01` through `text-05`, `text-inverted-XX`
+- **Backgrounds:** `background-neutral-XX`, `background-tint-XX` (and inverted variants)
+- **Borders:** `border-01` through `border-05`, `border-inverted-XX`
+- **Actions:** `action-link-XX`, `action-danger-XX`
+- **Status:** `status-info-XX`, `status-success-XX`, `status-warning-XX`, `status-error-XX`
+- **Theme:** `theme-primary-XX`, `theme-red-XX`, `theme-blue-XX`, etc.
+
+```typescript
+// ✅ Good - Use custom Onyx color classes
+<div className="bg-background-neutral-01 border border-border-02" />
+<div className="bg-background-tint-02 border border-border-01" />
+<div className="bg-status-success-01" />
+<div className="bg-action-link-01" />
+<div className="bg-theme-primary-05" />
+
+// ❌ Bad - Do NOT use standard Tailwind colors
+<div className="bg-gray-100 border border-gray-300 text-gray-600" />
+<div className="bg-white border border-slate-200" />
+<div className="bg-green-100 text-green-700" />
+<div className="bg-blue-100 text-blue-600" />
+<div className="bg-indigo-500" />
+```
+
+### 12. Data Fetching
+
+**Prefer using `useSWR` for data fetching. Data should generally be fetched on the client side. Components that need data should display a loader / placeholder while waiting for that data. Prefer loading data within the component that needs it rather than at the top level and passing it down.**
+
+**Reason:** Client side fetching allows us to load the skeleton of the page without waiting for data to load, leading to a snappier UX. Loading data where needed reduces dependencies between a component and its parent component(s).
+
 ## Database & Migrations
 
 ### Running Migrations
@@ -295,14 +575,6 @@ will be tailing their logs to this file.
 - Token management and rate limiting
 - Custom prompts and agent actions
 
-## UI/UX Patterns
-
-- Tailwind CSS with design system in `web/src/components/ui/`
-- Radix UI and Headless UI for accessible components
-- SWR for data fetching and caching
-- Form validation with react-hook-form
-- Error handling with popup notifications
-
 ## Creating a Plan
 When creating a plan in the `plans` directory, make sure to include at least these elements:
 

CLAUDE.md.template

@@ -7,7 +7,7 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co
 - If you run into any missing python dependency errors, try running your command with `source .venv/bin/activate` \
 to assume the python venv.
 - To make tests work, check the `.env` file at the root of the project to find an OpenAI key.
-- If using `playwright` to explore the frontend, you can usually log in with username `a@test.com` and password
+- If using `playwright` to explore the frontend, you can usually log in with username `a@example.com` and password
 `a`. The app can be accessed at `http://localhost:3000`.
 - You should assume that all Onyx services are running. To verify, you can check the `backend/log` directory to
 make sure we see logs coming out from the relevant service.
@@ -184,6 +184,286 @@ web/
 └── src/lib/                     # Utilities & business logic
 ```
 
+## Frontend Standards
+
+### 1. Import Standards
+
+**Always use absolute imports with the `@` prefix.**
+
+**Reason:** Moving files around becomes easier since you don't also have to update those import statements. This makes modifications to the codebase much nicer.
+
+```typescript
+// ✅ Good
+import { Button } from "@/components/ui/button";
+import { useAuth } from "@/hooks/useAuth";
+import { Text } from "@/refresh-components/texts/Text";
+
+// ❌ Bad
+import { Button } from "../../../components/ui/button";
+import { useAuth } from "./hooks/useAuth";
+```
+
+### 2. React Component Functions
+
+**Prefer regular functions over arrow functions for React components.**
+
+**Reason:** Functions just become easier to read.
+
+```typescript
+// ✅ Good
+function UserProfile({ userId }: UserProfileProps) {
+  return <div>User Profile</div>
+}
+
+// ❌ Bad
+const UserProfile = ({ userId }: UserProfileProps) => {
+  return <div>User Profile</div>
+}
+```
+
+### 3. Props Interface Extraction
+
+**Extract prop types into their own interface definitions.**
+
+**Reason:** Functions just become easier to read.
+
+```typescript
+// ✅ Good
+interface UserCardProps {
+  user: User
+  showActions?: boolean
+  onEdit?: (userId: string) => void
+}
+
+function UserCard({ user, showActions = false, onEdit }: UserCardProps) {
+  return <div>User Card</div>
+}
+
+// ❌ Bad
+function UserCard({
+  user,
+  showActions = false,
+  onEdit
+}: {
+  user: User
+  showActions?: boolean
+  onEdit?: (userId: string) => void
+}) {
+  return <div>User Card</div>
+}
+```
+
+### 4. Spacing Guidelines
+
+**Prefer padding over margins for spacing.**
+
+**Reason:** We want to consolidate usage to paddings instead of margins.
+
+```typescript
+// ✅ Good
+<div className="p-4 space-y-2">
+  <div className="p-2">Content</div>
+</div>
+
+// ❌ Bad
+<div className="m-4 space-y-2">
+  <div className="m-2">Content</div>
+</div>
+```
+
+### 5. Tailwind Dark Mode
+
+**Strictly forbid using the `dark:` modifier in Tailwind classes, except for logo icon handling.**
+
+**Reason:** The `colors.css` file already, VERY CAREFULLY, defines what the exact opposite colour of each light-mode colour is. Overriding this behaviour is VERY bad and will lead to horrible UI breakages.
+
+**Exception:** The `createLogoIcon` helper in `web/src/components/icons/icons.tsx` uses `dark:` modifiers (`dark:invert`, `dark:hidden`, `dark:block`) to handle third-party logo icons that cannot automatically adapt through `colors.css`. This is the ONLY acceptable use of dark mode modifiers.
+
+```typescript
+// ✅ Good - Standard components use `tailwind-themes/tailwind.config.js` / `src/app/css/colors.css`
+<div className="bg-background-neutral-03 text-text-02">
+  Content
+</div>
+
+// ✅ Good - Logo icons with dark mode handling via createLogoIcon
+export const GithubIcon = createLogoIcon(githubLightIcon, {
+  monochromatic: true,  // Will apply dark:invert internally
+});
+
+export const GitbookIcon = createLogoIcon(gitbookLightIcon, {
+  darkSrc: gitbookDarkIcon,  // Will use dark:hidden/dark:block internally
+});
+
+// ❌ Bad - Manual dark mode overrides
+<div className="bg-white dark:bg-black text-black dark:text-white">
+  Content
+</div>
+```
+
+### 6. Class Name Utilities
+
+**Use the `cn` utility instead of raw string formatting for classNames.**
+
+**Reason:** `cn`s are easier to read. They also allow for more complex types (i.e., string-arrays) to get formatted properly (it flattens each element in that string array down). As a result, it can allow things such as conditionals (i.e., `myCondition && "some-tailwind-class"`, which evaluates to `false` when `myCondition` is `false`) to get filtered out.
+
+```typescript
+import { cn } from '@/lib/utils'
+
+// ✅ Good
+<div className={cn(
+  'base-class',
+  isActive && 'active-class',
+  className
+)}>
+  Content
+</div>
+
+// ❌ Bad
+<div className={`base-class ${isActive ? 'active-class' : ''} ${className}`}>
+  Content
+</div>
+```
+
+### 7. Custom Hooks Organization
+
+**Follow a "hook-per-file" layout. Each hook should live in its own file within `web/src/hooks`.**
+
+**Reason:** This is just a layout preference. Keeps code clean.
+
+```typescript
+// web/src/hooks/useUserData.ts
+export function useUserData(userId: string) {
+  // hook implementation
+}
+
+// web/src/hooks/useLocalStorage.ts
+export function useLocalStorage<T>(key: string, initialValue: T) {
+  // hook implementation
+}
+```
+
+### 8. Icon Usage
+
+**ONLY use icons from the `web/src/icons` directory. Do NOT use icons from `react-icons`, `lucide`, or other external libraries.**
+
+**Reason:** We have a very carefully curated selection of icons that match our Onyx guidelines. We do NOT want to muddy those up with different aesthetic stylings.
+
+```typescript
+// ✅ Good
+import SvgX from "@/icons/x";
+import SvgMoreHorizontal from "@/icons/more-horizontal";
+
+// ❌ Bad
+import { User } from "lucide-react";
+import { FiSearch } from "react-icons/fi";
+```
+
+**Missing Icons**: If an icon is needed but doesn't exist in the `web/src/icons` directory, import it from Figma using the Figma MCP tool and add it to the icons directory.
+If you need help with this step, reach out to `raunak@onyx.app`.
+
+### 9. Text Rendering
+
+**Prefer using the `refresh-components/texts/Text` component for all text rendering. Avoid "naked" text nodes.**
+
+**Reason:** The `Text` component is fully compliant with the stylings provided in Figma. It provides easy utilities to specify the text-colour and font-size in the form of flags. Super duper easy.
+
+```typescript
+// ✅ Good
+import { Text } from '@/refresh-components/texts/Text'
+
+function UserCard({ name }: { name: string }) {
+  return (
+    <Text
+      {/* The `text03` flag makes the text it renders to be coloured the 3rd-scale grey */}
+      text03
+      {/* The `mainAction` flag makes the text it renders to be "main-action" font + line-height + weightage, as described in the Figma */}
+      mainAction
+    >
+      {name}
+    </Text>
+  )
+}
+
+// ❌ Bad
+function UserCard({ name }: { name: string }) {
+  return (
+    <div>
+      <h2>{name}</h2>
+      <p>User details</p>
+    </div>
+  )
+}
+```
+
+### 10. Component Usage
+
+**Heavily avoid raw HTML input components. Always use components from the `web/src/refresh-components` or `web/lib/opal/src` directory.**
+
+**Reason:** We've put in a lot of effort to unify the components that are rendered in the Onyx app. Using raw components breaks the entire UI of the application, and leaves it in a muddier state than before.
+
+```typescript
+// ✅ Good
+import Button from '@/refresh-components/buttons/Button'
+import InputTypeIn from '@/refresh-components/inputs/InputTypeIn'
+import SvgPlusCircle from '@/icons/plus-circle'
+
+function ContactForm() {
+  return (
+    <form>
+      <InputTypeIn placeholder="Search..." />
+      <Button type="submit" leftIcon={SvgPlusCircle}>Submit</Button>
+    </form>
+  )
+}
+
+// ❌ Bad
+function ContactForm() {
+  return (
+    <form>
+      <input placeholder="Name" />
+      <textarea placeholder="Message" />
+      <button type="submit">Submit</button>
+    </form>
+  )
+}
+```
+
+### 11. Colors
+
+**Always use custom overrides for colors and borders rather than built in Tailwind CSS colors. These overrides live in `web/tailwind-themes/tailwind.config.js`.**
+
+**Reason:** Our custom color system uses CSS variables that automatically handle dark mode and maintain design consistency across the app. Standard Tailwind colors bypass this system.
+
+**Available color categories:**
+- **Text:** `text-01` through `text-05`, `text-inverted-XX`
+- **Backgrounds:** `background-neutral-XX`, `background-tint-XX` (and inverted variants)
+- **Borders:** `border-01` through `border-05`, `border-inverted-XX`
+- **Actions:** `action-link-XX`, `action-danger-XX`
+- **Status:** `status-info-XX`, `status-success-XX`, `status-warning-XX`, `status-error-XX`
+- **Theme:** `theme-primary-XX`, `theme-red-XX`, `theme-blue-XX`, etc.
+
+```typescript
+// ✅ Good - Use custom Onyx color classes
+<div className="bg-background-neutral-01 border border-border-02" />
+<div className="bg-background-tint-02 border border-border-01" />
+<div className="bg-status-success-01" />
+<div className="bg-action-link-01" />
+<div className="bg-theme-primary-05" />
+
+// ❌ Bad - Do NOT use standard Tailwind colors
+<div className="bg-gray-100 border border-gray-300 text-gray-600" />
+<div className="bg-white border border-slate-200" />
+<div className="bg-green-100 text-green-700" />
+<div className="bg-blue-100 text-blue-600" />
+<div className="bg-indigo-500" />
+```
+
+### 12. Data Fetching
+
+**Prefer using `useSWR` for data fetching. Data should generally be fetched on the client side. Components that need data should display a loader / placeholder while waiting for that data. Prefer loading data within the component that needs it rather than at the top level and passing it down.**
+
+**Reason:** Client side fetching allows us to load the skeleton of the page without waiting for data to load, leading to a snappier UX. Loading data where needed reduces dependencies between a component and its parent component(s).
+
 ## Database & Migrations
 
 ### Running Migrations
@@ -300,14 +580,6 @@ will be tailing their logs to this file.
 - Token management and rate limiting
 - Custom prompts and agent actions
 
-## UI/UX Patterns
-
-- Tailwind CSS with design system in `web/src/components/ui/`
-- Radix UI and Headless UI for accessible components
-- SWR for data fetching and caching
-- Form validation with react-hook-form
-- Error handling with popup notifications
-
 ## Creating a Plan
 When creating a plan in the `plans` directory, make sure to include at least these elements:
 

backend/.dockerignore

@@ -15,3 +15,4 @@ build/
 dist/
 .coverage
 htmlcov/
+model_server/legacy/

backend/Dockerfile.model_server

@@ -13,23 +13,10 @@ RUN uv pip install --system --no-cache-dir --upgrade \
         -r /tmp/requirements.txt && \
     rm -rf ~/.cache/uv /tmp/*.txt
 
-# Stage for downloading tokenizers
-FROM base AS tokenizers
-RUN python -c "from transformers import AutoTokenizer; \
-AutoTokenizer.from_pretrained('distilbert-base-uncased'); \
-AutoTokenizer.from_pretrained('mixedbread-ai/mxbai-rerank-xsmall-v1');"
-
-# Stage for downloading Onyx models
-FROM base AS onyx-models
-RUN python -c "from huggingface_hub import snapshot_download; \
-snapshot_download(repo_id='onyx-dot-app/hybrid-intent-token-classifier'); \
-snapshot_download(repo_id='onyx-dot-app/information-content-model');"
-
-# Stage for downloading embedding and reranking models
+# Stage for downloading embedding models
 FROM base AS embedding-models
 RUN python -c "from huggingface_hub import snapshot_download; \
-snapshot_download('nomic-ai/nomic-embed-text-v1'); \
-snapshot_download('mixedbread-ai/mxbai-rerank-xsmall-v1');"
+snapshot_download('nomic-ai/nomic-embed-text-v1');"
 
 # Initialize SentenceTransformer to cache the custom architecture
 RUN python -c "from sentence_transformers import SentenceTransformer; \
@@ -54,8 +41,6 @@ RUN groupadd -g 1001 onyx && \
 # In case the user has volumes mounted to /app/.cache/huggingface that they've downloaded while
 # running Onyx, move the current contents of the cache folder to a temporary location to ensure
 # it's preserved in order to combine with the user's cache contents
-COPY --chown=onyx:onyx --from=tokenizers /app/.cache/huggingface /app/.cache/temp_huggingface
-COPY --chown=onyx:onyx --from=onyx-models /app/.cache/huggingface /app/.cache/temp_huggingface
 COPY --chown=onyx:onyx --from=embedding-models /app/.cache/huggingface /app/.cache/temp_huggingface
 
 WORKDIR /app

backend/alembic/env.py

@@ -39,7 +39,9 @@ config = context.config
 if config.config_file_name is not None and config.attributes.get(
     "configure_logger", True
 ):
-    fileConfig(config.config_file_name)
+    # disable_existing_loggers=False prevents breaking pytest's caplog fixture
+    # See: https://pytest-alembic.readthedocs.io/en/latest/setup.html#caplog-issues
+    fileConfig(config.config_file_name, disable_existing_loggers=False)
 
 target_metadata = [Base.metadata, ResultModelBase.metadata]
 
@@ -460,8 +462,49 @@ def run_migrations_offline() -> None:
 
 
 def run_migrations_online() -> None:
-    logger.info("run_migrations_online starting.")
-    asyncio.run(run_async_migrations())
+    """Run migrations in 'online' mode.
+
+    Supports pytest-alembic by checking for a pre-configured connection
+    in context.config.attributes["connection"]. If present, uses that
+    connection/engine directly instead of creating a new async engine.
+    """
+    # Check if pytest-alembic is providing a connection/engine
+    connectable = context.config.attributes.get("connection", None)
+
+    if connectable is not None:
+        # pytest-alembic is providing an engine - use it directly
+        logger.info("run_migrations_online starting (pytest-alembic mode).")
+
+        # For pytest-alembic, we use the default schema (public)
+        schema_name = context.config.attributes.get(
+            "schema_name", POSTGRES_DEFAULT_SCHEMA
+        )
+
+        # pytest-alembic passes an Engine, we need to get a connection from it
+        with connectable.connect() as connection:
+            # Set search path for the schema
+            connection.execute(text(f'SET search_path TO "{schema_name}"'))
+
+            context.configure(
+                connection=connection,
+                target_metadata=target_metadata,  # type: ignore
+                include_object=include_object,
+                version_table_schema=schema_name,
+                include_schemas=True,
+                compare_type=True,
+                compare_server_default=True,
+                script_location=config.get_main_option("script_location"),
+            )
+
+            with context.begin_transaction():
+                context.run_migrations()
+
+            # Commit the transaction to ensure changes are visible to next migration
+            connection.commit()
+    else:
+        # Normal operation - use async migrations
+        logger.info("run_migrations_online starting.")
+        asyncio.run(run_async_migrations())
 
 
 if context.is_offline_mode():

backend/alembic/versions/2b90f3af54b8_usage_limits.py

@@ -0,0 +1,46 @@
+"""usage_limits
+
+Revision ID: 2b90f3af54b8
+Revises: 9a0296d7421e
+Create Date: 2026-01-03 16:55:30.449692
+
+"""
+
+from alembic import op
+import sqlalchemy as sa
+
+
+# revision identifiers, used by Alembic.
+revision = "2b90f3af54b8"
+down_revision = "9a0296d7421e"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        "tenant_usage",
+        sa.Column("id", sa.Integer(), nullable=False),
+        sa.Column(
+            "window_start", sa.DateTime(timezone=True), nullable=False, index=True
+        ),
+        sa.Column("llm_cost_cents", sa.Float(), nullable=False, server_default="0.0"),
+        sa.Column("chunks_indexed", sa.Integer(), nullable=False, server_default="0"),
+        sa.Column("api_calls", sa.Integer(), nullable=False, server_default="0"),
+        sa.Column(
+            "non_streaming_api_calls", sa.Integer(), nullable=False, server_default="0"
+        ),
+        sa.Column(
+            "updated_at",
+            sa.DateTime(timezone=True),
+            server_default=sa.func.now(),
+            nullable=True,
+        ),
+        sa.PrimaryKeyConstraint("id"),
+        sa.UniqueConstraint("window_start", name="uq_tenant_usage_window"),
+    )
+
+
+def downgrade() -> None:
+    op.drop_index("ix_tenant_usage_window_start", table_name="tenant_usage")
+    op.drop_table("tenant_usage")

backend/alembic/versions/47a07e1a38f1_fix_invalid_model_configurations_state.py

@@ -11,7 +11,7 @@ from pydantic import BaseModel, ConfigDict
 import sqlalchemy as sa
 from sqlalchemy.dialects import postgresql
 
-from onyx.llm.llm_provider_options import (
+from onyx.llm.well_known_providers.llm_provider_options import (
     fetch_model_names_for_provider_as_set,
     fetch_visible_model_names_for_provider_as_set,
 )

backend/alembic/versions/4ea2c93919c1_add_type_to_credentials.py

@@ -62,6 +62,11 @@ def upgrade() -> None:
     )
     """
     )
+
+    # Drop the temporary table to avoid conflicts if migration runs again
+    # (e.g., during upgrade -> downgrade -> upgrade cycles in tests)
+    op.execute("DROP TABLE IF EXISTS temp_connector_credential")
+
     # If no exception was raised, alter the column
     op.alter_column("credential", "source", nullable=True)  # TODO modify
     # # ### end Alembic commands ###

backend/alembic/versions/5c3dca366b35_backend_driven_notification_details.py

@@ -0,0 +1,35 @@
+"""backend driven notification details
+
+Revision ID: 5c3dca366b35
+Revises: 9087b548dd69
+Create Date: 2026-01-06 16:03:11.413724
+
+"""
+
+from alembic import op
+import sqlalchemy as sa
+
+
+# revision identifiers, used by Alembic.
+revision = "5c3dca366b35"
+down_revision = "9087b548dd69"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        "notification",
+        sa.Column(
+            "title", sa.String(), nullable=False, server_default="New Notification"
+        ),
+    )
+    op.add_column(
+        "notification",
+        sa.Column("description", sa.String(), nullable=True, server_default=""),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("notification", "title")
+    op.drop_column("notification", "description")

backend/alembic/versions/699221885109_nullify_default_task_prompt.py

@@ -0,0 +1,75 @@
+"""nullify_default_task_prompt
+
+Revision ID: 699221885109
+Revises: 7e490836d179
+Create Date: 2025-12-30 10:00:00.000000
+
+"""
+
+from alembic import op
+import sqlalchemy as sa
+
+
+# revision identifiers, used by Alembic.
+revision = "699221885109"
+down_revision = "7e490836d179"
+branch_labels = None
+depends_on = None
+
+DEFAULT_PERSONA_ID = 0
+
+
+def upgrade() -> None:
+    # Make task_prompt column nullable
+    # Note: The model had nullable=True but the DB column was NOT NULL until this point
+    op.alter_column(
+        "persona",
+        "task_prompt",
+        nullable=True,
+    )
+
+    # Set task_prompt to NULL for the default persona
+    conn = op.get_bind()
+    conn.execute(
+        sa.text(
+            """
+            UPDATE persona
+            SET task_prompt = NULL
+            WHERE id = :persona_id
+            """
+        ),
+        {"persona_id": DEFAULT_PERSONA_ID},
+    )
+
+
+def downgrade() -> None:
+    # Restore task_prompt to empty string for the default persona
+    conn = op.get_bind()
+    conn.execute(
+        sa.text(
+            """
+            UPDATE persona
+            SET task_prompt = ''
+            WHERE id = :persona_id AND task_prompt IS NULL
+            """
+        ),
+        {"persona_id": DEFAULT_PERSONA_ID},
+    )
+
+    # Set any remaining NULL task_prompts to empty string before making non-nullable
+    conn.execute(
+        sa.text(
+            """
+            UPDATE persona
+            SET task_prompt = ''
+            WHERE task_prompt IS NULL
+            """
+        )
+    )
+
+    # Revert task_prompt column to not nullable
+    op.alter_column(
+        "persona",
+        "task_prompt",
+        nullable=False,
+    )

backend/alembic/versions/7206234e012a_add_image_generation_config_table.py

@@ -0,0 +1,54 @@
+"""add image generation config table
+
+Revision ID: 7206234e012a
+Revises: 699221885109
+Create Date: 2025-12-21 00:00:00.000000
+
+"""
+
+from alembic import op
+import sqlalchemy as sa
+
+
+# revision identifiers, used by Alembic.
+revision = "7206234e012a"
+down_revision = "699221885109"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        "image_generation_config",
+        sa.Column("image_provider_id", sa.String(), primary_key=True),
+        sa.Column("model_configuration_id", sa.Integer(), nullable=False),
+        sa.Column("is_default", sa.Boolean(), nullable=False),
+        sa.ForeignKeyConstraint(
+            ["model_configuration_id"],
+            ["model_configuration.id"],
+            ondelete="CASCADE",
+        ),
+    )
+    op.create_index(
+        "ix_image_generation_config_is_default",
+        "image_generation_config",
+        ["is_default"],
+        unique=False,
+    )
+    op.create_index(
+        "ix_image_generation_config_model_configuration_id",
+        "image_generation_config",
+        ["model_configuration_id"],
+        unique=False,
+    )
+
+
+def downgrade() -> None:
+    op.drop_index(
+        "ix_image_generation_config_model_configuration_id",
+        table_name="image_generation_config",
+    )
+    op.drop_index(
+        "ix_image_generation_config_is_default", table_name="image_generation_config"
+    )
+    op.drop_table("image_generation_config")

backend/alembic/versions/7a70b7664e37_add_model_configuration_table.py

@@ -10,7 +10,7 @@ from alembic import op
 import sqlalchemy as sa
 from sqlalchemy.dialects import postgresql
 
-from onyx.llm.llm_provider_options import (
+from onyx.llm.well_known_providers.llm_provider_options import (
     fetch_model_names_for_provider_as_set,
     fetch_visible_model_names_for_provider_as_set,
 )

backend/alembic/versions/7e490836d179_nullify_default_system_prompt.py

@@ -0,0 +1,80 @@
+"""nullify_default_system_prompt
+
+Revision ID: 7e490836d179
+Revises: c1d2e3f4a5b6
+Create Date: 2025-12-29 16:54:36.635574
+
+"""
+
+from alembic import op
+import sqlalchemy as sa
+
+
+# revision identifiers, used by Alembic.
+revision = "7e490836d179"
+down_revision = "c1d2e3f4a5b6"
+branch_labels = None
+depends_on = None
+
+
+# This is the default system prompt from the previous migration (87c52ec39f84)
+# ruff: noqa: E501, W605 start
+PREVIOUS_DEFAULT_SYSTEM_PROMPT = """
+You are a highly capable, thoughtful, and precise assistant. Your goal is to deeply understand the user's intent, ask clarifying questions when needed, think step-by-step through complex problems, provide clear and accurate answers, and proactively anticipate helpful follow-up information. Always prioritize being truthful, nuanced, insightful, and efficient.
+
+The current date is [[CURRENT_DATETIME]].[[CITATION_GUIDANCE]]
+
+# Response Style
+You use different text styles, bolding, emojis (sparingly), block quotes, and other formatting to make your responses more readable and engaging.
+You use proper Markdown and LaTeX to format your responses for math, scientific, and chemical formulas, symbols, etc.: '$$\\n[expression]\\n$$' for standalone cases and '\\( [expression] \\)' when inline.
+For code you prefer to use Markdown and specify the language.
+You can use horizontal rules (---) to separate sections of your responses.
+You can use Markdown tables to format your responses for data, lists, and other structured information.
+""".lstrip()
+# ruff: noqa: E501, W605 end
+
+
+def upgrade() -> None:
+    # Make system_prompt column nullable (model already has nullable=True but DB doesn't)
+    op.alter_column(
+        "persona",
+        "system_prompt",
+        nullable=True,
+    )
+
+    # Set system_prompt to NULL where it matches the previous default
+    conn = op.get_bind()
+    conn.execute(
+        sa.text(
+            """
+            UPDATE persona
+            SET system_prompt = NULL
+            WHERE system_prompt = :previous_default
+            """
+        ),
+        {"previous_default": PREVIOUS_DEFAULT_SYSTEM_PROMPT},
+    )
+
+
+def downgrade() -> None:
+    # Restore the default system prompt for personas that have NULL
+    # Note: This may restore the prompt to personas that originally had NULL
+    # before this migration, but there's no way to distinguish them
+    conn = op.get_bind()
+    conn.execute(
+        sa.text(
+            """
+            UPDATE persona
+            SET system_prompt = :previous_default
+            WHERE system_prompt IS NULL
+            """
+        ),
+        {"previous_default": PREVIOUS_DEFAULT_SYSTEM_PROMPT},
+    )
+
+    # Revert system_prompt column to not nullable
+    op.alter_column(
+        "persona",
+        "system_prompt",
+        nullable=False,
+    )

backend/alembic/versions/9087b548dd69_seed_default_image_gen_config.py

@@ -0,0 +1,136 @@
+"""seed_default_image_gen_config
+
+Revision ID: 9087b548dd69
+Revises: 2b90f3af54b8
+Create Date: 2026-01-05 00:00:00.000000
+
+"""
+
+from alembic import op
+import sqlalchemy as sa
+
+
+# revision identifiers, used by Alembic.
+revision = "9087b548dd69"
+down_revision = "2b90f3af54b8"
+branch_labels = None
+depends_on = None
+
+# Constants for default image generation config
+# Source: web/src/app/admin/configuration/image-generation/constants.ts
+IMAGE_PROVIDER_ID = "openai_gpt_image_1"
+MODEL_NAME = "gpt-image-1"
+PROVIDER_NAME = "openai"
+
+
+def upgrade() -> None:
+    conn = op.get_bind()
+
+    # Check if image_generation_config table already has records
+    existing_configs = (
+        conn.execute(sa.text("SELECT COUNT(*) FROM image_generation_config")).scalar()
+        or 0
+    )
+
+    if existing_configs > 0:
+        # Skip if configs already exist - user may have configured manually
+        return
+
+    # Find the first OpenAI LLM provider
+    openai_provider = conn.execute(
+        sa.text(
+            """
+            SELECT id, api_key
+            FROM llm_provider
+            WHERE provider = :provider
+            ORDER BY id
+            LIMIT 1
+            """
+        ),
+        {"provider": PROVIDER_NAME},
+    ).fetchone()
+
+    if not openai_provider:
+        # No OpenAI provider found - nothing to do
+        return
+
+    source_provider_id, api_key = openai_provider
+
+    # Create new LLM provider for image generation (clone only api_key)
+    result = conn.execute(
+        sa.text(
+            """
+            INSERT INTO llm_provider (
+                name, provider, api_key, api_base, api_version,
+                deployment_name, default_model_name, is_public,
+                is_default_provider, is_default_vision_provider, is_auto_mode
+            )
+            VALUES (
+                :name, :provider, :api_key, NULL, NULL,
+                NULL, :default_model_name, :is_public,
+                NULL, NULL, :is_auto_mode
+            )
+            RETURNING id
+            """
+        ),
+        {
+            "name": f"Image Gen - {IMAGE_PROVIDER_ID}",
+            "provider": PROVIDER_NAME,
+            "api_key": api_key,
+            "default_model_name": MODEL_NAME,
+            "is_public": True,
+            "is_auto_mode": False,
+        },
+    )
+    new_provider_id = result.scalar()
+
+    # Create model configuration
+    result = conn.execute(
+        sa.text(
+            """
+            INSERT INTO model_configuration (
+                llm_provider_id, name, is_visible, max_input_tokens,
+                supports_image_input, display_name
+            )
+            VALUES (
+                :llm_provider_id, :name, :is_visible, :max_input_tokens,
+                :supports_image_input, :display_name
+            )
+            RETURNING id
+            """
+        ),
+        {
+            "llm_provider_id": new_provider_id,
+            "name": MODEL_NAME,
+            "is_visible": True,
+            "max_input_tokens": None,
+            "supports_image_input": False,
+            "display_name": None,
+        },
+    )
+    model_config_id = result.scalar()
+
+    # Create image generation config
+    conn.execute(
+        sa.text(
+            """
+            INSERT INTO image_generation_config (
+                image_provider_id, model_configuration_id, is_default
+            )
+            VALUES (
+                :image_provider_id, :model_configuration_id, :is_default
+            )
+            """
+        ),
+        {
+            "image_provider_id": IMAGE_PROVIDER_ID,
+            "model_configuration_id": model_config_id,
+            "is_default": True,
+        },
+    )
+
+
+def downgrade() -> None:
+    # We don't remove the config on downgrade since it's safe to keep around
+    # If we upgrade again, it will be a no-op due to the existing records check
+    pass

backend/alembic/versions/9a0296d7421e_add_is_auto_mode_to_llm_provider.py

@@ -0,0 +1,33 @@
+"""add_is_auto_mode_to_llm_provider
+
+Revision ID: 9a0296d7421e
+Revises: 7206234e012a
+Create Date: 2025-12-17 18:14:29.620981
+
+"""
+
+from alembic import op
+import sqlalchemy as sa
+
+
+# revision identifiers, used by Alembic.
+revision = "9a0296d7421e"
+down_revision = "7206234e012a"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        "llm_provider",
+        sa.Column(
+            "is_auto_mode",
+            sa.Boolean(),
+            nullable=False,
+            server_default="false",
+        ),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("llm_provider", "is_auto_mode")

backend/alembic/versions/9b66d3156fc6_user_file_schema_additions.py

@@ -234,6 +234,8 @@ def downgrade() -> None:
         if "instructions" in columns:
             op.drop_column("user_project", "instructions")
         op.execute("ALTER TABLE user_project RENAME TO user_folder")
+        # Update NULL descriptions to empty string before setting NOT NULL constraint
+        op.execute("UPDATE user_folder SET description = '' WHERE description IS NULL")
         op.alter_column("user_folder", "description", nullable=False)
         logger.info("Renamed user_project back to user_folder")
 

backend/alembic/versions/a3c1a7904cd0_remove_userfile_related_deprecated_.py

@@ -0,0 +1,39 @@
+"""remove userfile related deprecated fields
+
+Revision ID: a3c1a7904cd0
+Revises: 5c3dca366b35
+Create Date: 2026-01-06 13:00:30.634396
+
+"""
+
+from alembic import op
+import sqlalchemy as sa
+
+# revision identifiers, used by Alembic.
+revision = "a3c1a7904cd0"
+down_revision = "5c3dca366b35"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.drop_column("user_file", "document_id")
+    op.drop_column("user_file", "document_id_migrated")
+    op.drop_column("connector_credential_pair", "is_user_file")
+
+
+def downgrade() -> None:
+    op.add_column(
+        "connector_credential_pair",
+        sa.Column("is_user_file", sa.Boolean(), nullable=False, server_default="false"),
+    )
+    op.add_column(
+        "user_file",
+        sa.Column("document_id", sa.String(), nullable=True),
+    )
+    op.add_column(
+        "user_file",
+        sa.Column(
+            "document_id_migrated", sa.Boolean(), nullable=False, server_default="true"
+        ),
+    )

backend/alembic/versions/a852cbe15577_new_chat_history.py

@@ -280,6 +280,14 @@ def downgrade() -> None:
     op.add_column(
         "chat_message", sa.Column("alternate_assistant_id", sa.Integer(), nullable=True)
     )
+    # Recreate the FK constraint that was implicitly dropped when the column was dropped
+    op.create_foreign_key(
+        "fk_chat_message_persona",
+        "chat_message",
+        "persona",
+        ["alternate_assistant_id"],
+        ["id"],
+    )
     op.add_column(
         "chat_message", sa.Column("rephrased_query", sa.Text(), nullable=True)
     )

backend/alembic_tenants/env.py

@@ -20,7 +20,9 @@ config = context.config
 if config.config_file_name is not None and config.attributes.get(
     "configure_logger", True
 ):
-    fileConfig(config.config_file_name)
+    # disable_existing_loggers=False prevents breaking pytest's caplog fixture
+    # See: https://pytest-alembic.readthedocs.io/en/latest/setup.html#caplog-issues
+    fileConfig(config.config_file_name, disable_existing_loggers=False)
 
 # add your model's MetaData object here
 # for 'autogenerate' support
@@ -108,9 +110,24 @@ async def run_async_migrations() -> None:
 
 
 def run_migrations_online() -> None:
-    """Run migrations in 'online' mode."""
+    """Run migrations in 'online' mode.
 
-    asyncio.run(run_async_migrations())
+    Supports pytest-alembic by checking for a pre-configured connection
+    in context.config.attributes["connection"]. If present, uses that
+    connection/engine directly instead of creating a new async engine.
+    """
+    # Check if pytest-alembic is providing a connection/engine
+    connectable = context.config.attributes.get("connection", None)
+
+    if connectable is not None:
+        # pytest-alembic is providing an engine - use it directly
+        with connectable.connect() as connection:
+            do_run_migrations(connection)
+            # Commit to ensure changes are visible to next migration
+            connection.commit()
+    else:
+        # Normal operation - use async migrations
+        asyncio.run(run_async_migrations())
 
 
 if context.is_offline_mode():

backend/ee/onyx/configs/app_configs.py

@@ -111,10 +111,6 @@ CHECK_TTL_MANAGEMENT_TASK_FREQUENCY_IN_HOURS = float(
 STRIPE_SECRET_KEY = os.environ.get("STRIPE_SECRET_KEY")
 STRIPE_PRICE_ID = os.environ.get("STRIPE_PRICE")
 
-OPENAI_DEFAULT_API_KEY = os.environ.get("OPENAI_DEFAULT_API_KEY")
-ANTHROPIC_DEFAULT_API_KEY = os.environ.get("ANTHROPIC_DEFAULT_API_KEY")
-COHERE_DEFAULT_API_KEY = os.environ.get("COHERE_DEFAULT_API_KEY")
-
 # JWT Public Key URL
 JWT_PUBLIC_KEY_URL: str | None = os.getenv("JWT_PUBLIC_KEY_URL", None)
 

backend/ee/onyx/db/persona.py

@@ -34,6 +34,7 @@ def make_persona_private(
                 create_notification(
                     user_id=user_id,
                     notif_type=NotificationType.PERSONA_SHARED,
+                    title="A new agent was shared with you!",
                     db_session=db_session,
                     additional_data=PersonaSharedNotificationData(
                         persona_id=persona_id,

backend/ee/onyx/server/analytics/api.py

@@ -21,8 +21,9 @@ from onyx.auth.users import current_admin_user
 from onyx.auth.users import current_user
 from onyx.db.engine.sql_engine import get_session
 from onyx.db.models import User
+from onyx.server.utils import PUBLIC_API_TAGS
 
-router = APIRouter(prefix="/analytics")
+router = APIRouter(prefix="/analytics", tags=PUBLIC_API_TAGS)
 
 
 _DEFAULT_LOOKBACK_DAYS = 30

backend/ee/onyx/server/enterprise_settings/models.py

@@ -1,3 +1,4 @@
+from enum import Enum
 from typing import Any
 from typing import List
 
@@ -23,6 +24,12 @@ class NavigationItem(BaseModel):
         return instance
 
 
+class LogoDisplayStyle(str, Enum):
+    LOGO_AND_NAME = "logo_and_name"
+    LOGO_ONLY = "logo_only"
+    NAME_ONLY = "name_only"
+
+
 class EnterpriseSettings(BaseModel):
     """General settings that only apply to the Enterprise Edition of Onyx
 
@@ -31,6 +38,7 @@ class EnterpriseSettings(BaseModel):
     application_name: str | None = None
     use_custom_logo: bool = False
     use_custom_logotype: bool = False
+    logo_display_style: LogoDisplayStyle | None = None
 
     # custom navigation
     custom_nav_items: List[NavigationItem] = Field(default_factory=list)
@@ -42,6 +50,9 @@ class EnterpriseSettings(BaseModel):
     custom_popup_header: str | None = None
     custom_popup_content: str | None = None
     enable_consent_screen: bool | None = None
+    consent_screen_prompt: str | None = None
+    show_first_visit_notice: bool | None = None
+    custom_greeting_message: str | None = None
 
     def check_validity(self) -> None:
         return

backend/ee/onyx/server/query_and_chat/chat_backend.py

@@ -106,7 +106,6 @@ def handle_simplified_chat_message(
         new_msg_req=full_chat_msg_info,
         user=user,
         db_session=db_session,
-        enforce_chat_session_id_for_search_docs=False,
     )
 
     return gather_stream(packets)
@@ -210,7 +209,6 @@ def handle_send_message_simple_with_history(
         new_msg_req=full_chat_msg_info,
         user=user,
         db_session=db_session,
-        enforce_chat_session_id_for_search_docs=False,
     )
 
     return gather_stream(packets)

backend/ee/onyx/server/query_history/api.py

@@ -48,6 +48,7 @@ from onyx.file_store.file_store import get_default_file_store
 from onyx.server.documents.models import PaginatedReturn
 from onyx.server.query_and_chat.models import ChatSessionDetails
 from onyx.server.query_and_chat.models import ChatSessionsResponse
+from onyx.server.utils import PUBLIC_API_TAGS
 from onyx.utils.threadpool_concurrency import parallel_yield
 from shared_configs.contextvars import get_current_tenant_id
 
@@ -294,7 +295,7 @@ def list_all_query_history_exports(
         )
 
 
-@router.post("/admin/query-history/start-export")
+@router.post("/admin/query-history/start-export", tags=PUBLIC_API_TAGS)
 def start_query_history_export(
     _: User | None = Depends(current_admin_user),
     db_session: Session = Depends(get_session),
@@ -340,7 +341,7 @@ def start_query_history_export(
     return {"request_id": task_id}
 
 
-@router.get("/admin/query-history/export-status")
+@router.get("/admin/query-history/export-status", tags=PUBLIC_API_TAGS)
 def get_query_history_export_status(
     request_id: str,
     _: User | None = Depends(current_admin_user),
@@ -374,7 +375,7 @@ def get_query_history_export_status(
     return {"status": TaskStatus.SUCCESS}
 
 
-@router.get("/admin/query-history/download")
+@router.get("/admin/query-history/download", tags=PUBLIC_API_TAGS)
 def download_query_history_csv(
     request_id: str,
     _: User | None = Depends(current_admin_user),

backend/ee/onyx/server/tenant_usage_limits.py

@@ -0,0 +1,92 @@
+"""Tenant-specific usage limit overrides from the control plane (EE version)."""
+
+import requests
+
+from ee.onyx.server.tenants.access import generate_data_plane_token
+from onyx.configs.app_configs import CONTROL_PLANE_API_BASE_URL
+from onyx.server.tenant_usage_limits import TenantUsageLimitOverrides
+from onyx.utils.logger import setup_logger
+
+logger = setup_logger()
+
+
+# In-memory storage for tenant overrides (populated at startup)
+_tenant_usage_limit_overrides: dict[str, TenantUsageLimitOverrides] | None = None
+
+
+def fetch_usage_limit_overrides() -> dict[str, TenantUsageLimitOverrides]:
+    """
+    Fetch tenant-specific usage limit overrides from the control plane.
+
+    Returns:
+        Dictionary mapping tenant_id to their specific limit overrides.
+        Returns empty dict on any error (falls back to defaults).
+    """
+    try:
+        token = generate_data_plane_token()
+        headers = {
+            "Authorization": f"Bearer {token}",
+            "Content-Type": "application/json",
+        }
+        url = f"{CONTROL_PLANE_API_BASE_URL}/usage-limit-overrides"
+        response = requests.get(url, headers=headers, timeout=30)
+        response.raise_for_status()
+
+        tenant_overrides = response.json()
+
+        # Parse each tenant's overrides
+        result: dict[str, TenantUsageLimitOverrides] = {}
+        for override_data in tenant_overrides:
+            tenant_id = override_data["tenant_id"]
+            try:
+                result[tenant_id] = TenantUsageLimitOverrides(**override_data)
+            except Exception as e:
+                logger.warning(
+                    f"Failed to parse usage limit overrides for tenant {tenant_id}: {e}"
+                )
+
+        return result
+
+    except requests.exceptions.RequestException as e:
+        logger.warning(f"Failed to fetch usage limit overrides from control plane: {e}")
+        return {}
+    except Exception as e:
+        logger.error(f"Error parsing usage limit overrides: {e}")
+        return {}
+
+
+def load_usage_limit_overrides() -> dict[str, TenantUsageLimitOverrides]:
+    """
+    Load tenant usage limit overrides from the control plane.
+
+    Called at server startup to populate the in-memory cache.
+    """
+    global _tenant_usage_limit_overrides
+
+    logger.info("Loading tenant usage limit overrides from control plane...")
+    overrides = fetch_usage_limit_overrides()
+    _tenant_usage_limit_overrides = overrides
+
+    if overrides:
+        logger.info(f"Loaded usage limit overrides for {len(overrides)} tenants")
+    else:
+        logger.info("No tenant-specific usage limit overrides found")
+    return overrides
+
+
+def get_tenant_usage_limit_overrides(
+    tenant_id: str,
+) -> TenantUsageLimitOverrides | None:
+    """
+    Get the usage limit overrides for a specific tenant.
+
+    Args:
+        tenant_id: The tenant ID to look up
+
+    Returns:
+        TenantUsageLimitOverrides if the tenant has overrides, None otherwise.
+    """
+    global _tenant_usage_limit_overrides
+    if _tenant_usage_limit_overrides is None:
+        _tenant_usage_limit_overrides = load_usage_limit_overrides()
+    return _tenant_usage_limit_overrides.get(tenant_id)

backend/ee/onyx/server/tenants/provisioning.py

@@ -1,5 +1,4 @@
 import asyncio
-import logging
 import uuid
 
 import aiohttp  # Async HTTP client
@@ -10,10 +9,7 @@ from fastapi import Request
 from sqlalchemy import select
 from sqlalchemy.orm import Session
 
-from ee.onyx.configs.app_configs import ANTHROPIC_DEFAULT_API_KEY
-from ee.onyx.configs.app_configs import COHERE_DEFAULT_API_KEY
 from ee.onyx.configs.app_configs import HUBSPOT_TRACKING_URL
-from ee.onyx.configs.app_configs import OPENAI_DEFAULT_API_KEY
 from ee.onyx.server.tenants.access import generate_data_plane_token
 from ee.onyx.server.tenants.models import TenantByDomainResponse
 from ee.onyx.server.tenants.models import TenantCreationPayload
@@ -25,11 +21,18 @@ from ee.onyx.server.tenants.user_mapping import add_users_to_tenant
 from ee.onyx.server.tenants.user_mapping import get_tenant_id_for_email
 from ee.onyx.server.tenants.user_mapping import user_owns_a_tenant
 from onyx.auth.users import exceptions
+from onyx.configs.app_configs import ANTHROPIC_DEFAULT_API_KEY
+from onyx.configs.app_configs import COHERE_DEFAULT_API_KEY
 from onyx.configs.app_configs import CONTROL_PLANE_API_BASE_URL
 from onyx.configs.app_configs import DEV_MODE
+from onyx.configs.app_configs import OPENAI_DEFAULT_API_KEY
+from onyx.configs.app_configs import OPENROUTER_DEFAULT_API_KEY
+from onyx.configs.app_configs import VERTEXAI_DEFAULT_CREDENTIALS
+from onyx.configs.app_configs import VERTEXAI_DEFAULT_LOCATION
 from onyx.configs.constants import MilestoneRecordType
 from onyx.db.engine.sql_engine import get_session_with_shared_schema
 from onyx.db.engine.sql_engine import get_session_with_tenant
+from onyx.db.image_generation import create_default_image_gen_config_from_api_key
 from onyx.db.llm import update_default_provider
 from onyx.db.llm import upsert_cloud_embedding_provider
 from onyx.db.llm import upsert_llm_provider
@@ -37,14 +40,24 @@ from onyx.db.models import AvailableTenant
 from onyx.db.models import IndexModelStatus
 from onyx.db.models import SearchSettings
 from onyx.db.models import UserTenantMapping
-from onyx.llm.llm_provider_options import ANTHROPIC_PROVIDER_NAME
-from onyx.llm.llm_provider_options import get_anthropic_model_names
-from onyx.llm.llm_provider_options import get_openai_model_names
-from onyx.llm.llm_provider_options import OPENAI_PROVIDER_NAME
+from onyx.llm.well_known_providers.auto_update_models import LLMRecommendations
+from onyx.llm.well_known_providers.constants import ANTHROPIC_PROVIDER_NAME
+from onyx.llm.well_known_providers.constants import OPENAI_PROVIDER_NAME
+from onyx.llm.well_known_providers.constants import OPENROUTER_PROVIDER_NAME
+from onyx.llm.well_known_providers.constants import VERTEX_CREDENTIALS_FILE_KWARG
+from onyx.llm.well_known_providers.constants import VERTEX_LOCATION_KWARG
+from onyx.llm.well_known_providers.constants import VERTEXAI_PROVIDER_NAME
+from onyx.llm.well_known_providers.llm_provider_options import (
+    get_recommendations,
+)
+from onyx.llm.well_known_providers.llm_provider_options import (
+    model_configurations_for_provider,
+)
 from onyx.server.manage.embedding.models import CloudEmbeddingProviderCreationRequest
 from onyx.server.manage.llm.models import LLMProviderUpsertRequest
 from onyx.server.manage.llm.models import ModelConfigurationUpsertRequest
 from onyx.setup import setup_onyx
+from onyx.utils.logger import setup_logger
 from onyx.utils.telemetry import mt_cloud_telemetry
 from shared_configs.configs import MULTI_TENANT
 from shared_configs.configs import POSTGRES_DEFAULT_SCHEMA
@@ -53,7 +66,7 @@ from shared_configs.contextvars import CURRENT_TENANT_ID_CONTEXTVAR
 from shared_configs.enums import EmbeddingProvider
 
 
-logger = logging.getLogger(__name__)
+logger = setup_logger()
 
 
 async def get_or_provision_tenant(
@@ -262,59 +275,173 @@ async def rollback_tenant_provisioning(tenant_id: str) -> None:
         logger.info(f"Tenant rollback completed successfully for tenant {tenant_id}")
 
 
-def configure_default_api_keys(db_session: Session) -> None:
-    if ANTHROPIC_DEFAULT_API_KEY:
-        anthropic_provider = LLMProviderUpsertRequest(
-            name="Anthropic",
-            provider=ANTHROPIC_PROVIDER_NAME,
-            api_key=ANTHROPIC_DEFAULT_API_KEY,
-            default_model_name="claude-3-7-sonnet-20250219",
-            model_configurations=[
-                ModelConfigurationUpsertRequest(
-                    name=name,
-                    is_visible=False,
-                    max_input_tokens=None,
-                )
-                for name in get_anthropic_model_names()
-            ],
-            api_key_changed=True,
-        )
-        try:
-            full_provider = upsert_llm_provider(anthropic_provider, db_session)
-            update_default_provider(full_provider.id, db_session)
-        except Exception as e:
-            logger.error(f"Failed to configure Anthropic provider: {e}")
-    else:
-        logger.error(
-            "ANTHROPIC_DEFAULT_API_KEY not set, skipping Anthropic provider configuration"
+def _build_model_configuration_upsert_requests(
+    provider_name: str,
+    recommendations: LLMRecommendations,
+) -> list[ModelConfigurationUpsertRequest]:
+    model_configurations = model_configurations_for_provider(
+        provider_name, recommendations
+    )
+    return [
+        ModelConfigurationUpsertRequest(
+            name=model_configuration.name,
+            is_visible=model_configuration.is_visible,
+            max_input_tokens=model_configuration.max_input_tokens,
+            supports_image_input=model_configuration.supports_image_input,
         )
+        for model_configuration in model_configurations
+    ]
 
+
+def configure_default_api_keys(db_session: Session) -> None:
+    """Configure default LLM providers using recommended-models.json for model selection."""
+    # Load recommendations from JSON config
+    recommendations = get_recommendations()
+
+    has_set_default_provider = False
+
+    def _upsert(request: LLMProviderUpsertRequest) -> None:
+        nonlocal has_set_default_provider
+        try:
+            provider = upsert_llm_provider(request, db_session)
+            if not has_set_default_provider:
+                update_default_provider(provider.id, db_session)
+                has_set_default_provider = True
+        except Exception as e:
+            logger.error(f"Failed to configure {request.provider} provider: {e}")
+
+    # Configure OpenAI provider
     if OPENAI_DEFAULT_API_KEY:
+        default_model = recommendations.get_default_model(OPENAI_PROVIDER_NAME)
+        if default_model is None:
+            logger.error(
+                f"No default model found for {OPENAI_PROVIDER_NAME} in recommendations"
+            )
+        default_model_name = default_model.name if default_model else "gpt-5.2"
+
         openai_provider = LLMProviderUpsertRequest(
             name="OpenAI",
             provider=OPENAI_PROVIDER_NAME,
             api_key=OPENAI_DEFAULT_API_KEY,
-            default_model_name="gpt-4o",
-            model_configurations=[
-                ModelConfigurationUpsertRequest(
-                    name=model_name,
-                    is_visible=False,
-                    max_input_tokens=None,
-                )
-                for model_name in get_openai_model_names()
-            ],
+            default_model_name=default_model_name,
+            model_configurations=_build_model_configuration_upsert_requests(
+                OPENAI_PROVIDER_NAME, recommendations
+            ),
             api_key_changed=True,
+            is_auto_mode=True,
         )
+        _upsert(openai_provider)
+
+        # Create default image generation config using the OpenAI API key
         try:
-            full_provider = upsert_llm_provider(openai_provider, db_session)
-            update_default_provider(full_provider.id, db_session)
+            create_default_image_gen_config_from_api_key(
+                db_session, OPENAI_DEFAULT_API_KEY
+            )
         except Exception as e:
-            logger.error(f"Failed to configure OpenAI provider: {e}")
+            logger.error(f"Failed to create default image gen config: {e}")
     else:
-        logger.error(
+        logger.info(
             "OPENAI_DEFAULT_API_KEY not set, skipping OpenAI provider configuration"
         )
 
+    # Configure Anthropic provider
+    if ANTHROPIC_DEFAULT_API_KEY:
+        default_model = recommendations.get_default_model(ANTHROPIC_PROVIDER_NAME)
+        if default_model is None:
+            logger.error(
+                f"No default model found for {ANTHROPIC_PROVIDER_NAME} in recommendations"
+            )
+        default_model_name = (
+            default_model.name if default_model else "claude-sonnet-4-5"
+        )
+
+        anthropic_provider = LLMProviderUpsertRequest(
+            name="Anthropic",
+            provider=ANTHROPIC_PROVIDER_NAME,
+            api_key=ANTHROPIC_DEFAULT_API_KEY,
+            default_model_name=default_model_name,
+            model_configurations=_build_model_configuration_upsert_requests(
+                ANTHROPIC_PROVIDER_NAME, recommendations
+            ),
+            api_key_changed=True,
+            is_auto_mode=True,
+        )
+        _upsert(anthropic_provider)
+    else:
+        logger.info(
+            "ANTHROPIC_DEFAULT_API_KEY not set, skipping Anthropic provider configuration"
+        )
+
+    # Configure Vertex AI provider
+    if VERTEXAI_DEFAULT_CREDENTIALS:
+        default_model = recommendations.get_default_model(VERTEXAI_PROVIDER_NAME)
+        if default_model is None:
+            logger.error(
+                f"No default model found for {VERTEXAI_PROVIDER_NAME} in recommendations"
+            )
+        default_model_name = default_model.name if default_model else "gemini-2.5-pro"
+
+        # Vertex AI uses custom_config for credentials and location
+        custom_config = {
+            VERTEX_CREDENTIALS_FILE_KWARG: VERTEXAI_DEFAULT_CREDENTIALS,
+            VERTEX_LOCATION_KWARG: VERTEXAI_DEFAULT_LOCATION,
+        }
+
+        vertexai_provider = LLMProviderUpsertRequest(
+            name="Google Vertex AI",
+            provider=VERTEXAI_PROVIDER_NAME,
+            custom_config=custom_config,
+            default_model_name=default_model_name,
+            model_configurations=_build_model_configuration_upsert_requests(
+                VERTEXAI_PROVIDER_NAME, recommendations
+            ),
+            api_key_changed=True,
+            is_auto_mode=True,
+        )
+        _upsert(vertexai_provider)
+    else:
+        logger.info(
+            "VERTEXAI_DEFAULT_CREDENTIALS not set, skipping Vertex AI provider configuration"
+        )
+
+    # Configure OpenRouter provider
+    if OPENROUTER_DEFAULT_API_KEY:
+        default_model = recommendations.get_default_model(OPENROUTER_PROVIDER_NAME)
+        if default_model is None:
+            logger.error(
+                f"No default model found for {OPENROUTER_PROVIDER_NAME} in recommendations"
+            )
+        default_model_name = default_model.name if default_model else "z-ai/glm-4.7"
+
+        # For OpenRouter, we use the visible models from recommendations as model_configurations
+        # since OpenRouter models are dynamic (fetched from their API)
+        visible_models = recommendations.get_visible_models(OPENROUTER_PROVIDER_NAME)
+        model_configurations = [
+            ModelConfigurationUpsertRequest(
+                name=model.name,
+                is_visible=True,
+                max_input_tokens=None,
+                display_name=model.display_name,
+            )
+            for model in visible_models
+        ]
+
+        openrouter_provider = LLMProviderUpsertRequest(
+            name="OpenRouter",
+            provider=OPENROUTER_PROVIDER_NAME,
+            api_key=OPENROUTER_DEFAULT_API_KEY,
+            default_model_name=default_model_name,
+            model_configurations=model_configurations,
+            api_key_changed=True,
+            is_auto_mode=True,
+        )
+        _upsert(openrouter_provider)
+    else:
+        logger.info(
+            "OPENROUTER_DEFAULT_API_KEY not set, skipping OpenRouter provider configuration"
+        )
+
+    # Configure Cohere embedding provider
     if COHERE_DEFAULT_API_KEY:
         cloud_embedding_provider = CloudEmbeddingProviderCreationRequest(
             provider_type=EmbeddingProvider.COHERE,

backend/ee/onyx/server/token_rate_limits/api.py

@@ -16,8 +16,9 @@ from onyx.db.token_limit import insert_user_token_rate_limit
 from onyx.server.query_and_chat.token_limit import any_rate_limit_exists
 from onyx.server.token_rate_limits.models import TokenRateLimitArgs
 from onyx.server.token_rate_limits.models import TokenRateLimitDisplay
+from onyx.server.utils import PUBLIC_API_TAGS
 
-router = APIRouter(prefix="/admin/token-rate-limits")
+router = APIRouter(prefix="/admin/token-rate-limits", tags=PUBLIC_API_TAGS)
 
 
 """

backend/ee/onyx/server/usage_limits.py

@@ -0,0 +1,47 @@
+"""EE Usage limits - trial detection via billing information."""
+
+from datetime import datetime
+from datetime import timezone
+
+from ee.onyx.server.tenants.billing import fetch_billing_information
+from ee.onyx.server.tenants.models import BillingInformation
+from ee.onyx.server.tenants.models import SubscriptionStatusResponse
+from onyx.utils.logger import setup_logger
+from shared_configs.configs import MULTI_TENANT
+
+logger = setup_logger()
+
+
+def is_tenant_on_trial(tenant_id: str) -> bool:
+    """
+    Determine if a tenant is currently on a trial subscription.
+
+    In multi-tenant mode, we fetch billing information from the control plane
+    to determine if the tenant has an active trial.
+    """
+    if not MULTI_TENANT:
+        return False
+
+    try:
+        billing_info = fetch_billing_information(tenant_id)
+
+        # If not subscribed at all, check if we have trial information
+        if isinstance(billing_info, SubscriptionStatusResponse):
+            # No subscription means they're likely on trial (new tenant)
+            return True
+
+        if isinstance(billing_info, BillingInformation):
+            # Check if trial is active
+            if billing_info.trial_end is not None:
+                now = datetime.now(timezone.utc)
+                # Trial active if trial_end is in the future
+                # and subscription status indicates trialing
+                if billing_info.trial_end > now and billing_info.status == "trialing":
+                    return True
+
+        return False
+
+    except Exception as e:
+        logger.warning(f"Failed to fetch billing info for trial check: {e}")
+        # Default to trial limits on error (more restrictive = safer)
+        return True

backend/ee/onyx/server/user_group/api.py

@@ -21,11 +21,12 @@ from onyx.auth.users import current_curator_or_admin_user
 from onyx.db.engine.sql_engine import get_session
 from onyx.db.models import User
 from onyx.db.models import UserRole
+from onyx.server.utils import PUBLIC_API_TAGS
 from onyx.utils.logger import setup_logger
 
 logger = setup_logger()
 
-router = APIRouter(prefix="/manage")
+router = APIRouter(prefix="/manage", tags=PUBLIC_API_TAGS)
 
 
 @router.get("/admin/user-group")

backend/model_server/constants.py

@@ -1,5 +1,4 @@
 MODEL_WARM_UP_STRING = "hi " * 512
-INFORMATION_CONTENT_MODEL_WARM_UP_STRING = "hi " * 16
 
 
 class GPUStatus:

backend/model_server/custom_models.py

@@ -1,562 +0,0 @@
-from typing import cast
-from typing import Optional
-from typing import TYPE_CHECKING
-
-import numpy as np
-import torch
-import torch.nn.functional as F
-from fastapi import APIRouter
-from huggingface_hub import snapshot_download
-
-from model_server.constants import INFORMATION_CONTENT_MODEL_WARM_UP_STRING
-from model_server.constants import MODEL_WARM_UP_STRING
-from model_server.onyx_torch_model import ConnectorClassifier
-from model_server.onyx_torch_model import HybridClassifier
-from model_server.utils import simple_log_function_time
-from onyx.utils.logger import setup_logger
-from shared_configs.configs import CONNECTOR_CLASSIFIER_MODEL_REPO
-from shared_configs.configs import CONNECTOR_CLASSIFIER_MODEL_TAG
-from shared_configs.configs import (
-    INDEXING_INFORMATION_CONTENT_CLASSIFICATION_CUTOFF_LENGTH,
-)
-from shared_configs.configs import INDEXING_INFORMATION_CONTENT_CLASSIFICATION_MAX
-from shared_configs.configs import INDEXING_INFORMATION_CONTENT_CLASSIFICATION_MIN
-from shared_configs.configs import (
-    INDEXING_INFORMATION_CONTENT_CLASSIFICATION_TEMPERATURE,
-)
-from shared_configs.configs import INDEXING_ONLY
-from shared_configs.configs import INFORMATION_CONTENT_MODEL_TAG
-from shared_configs.configs import INFORMATION_CONTENT_MODEL_VERSION
-from shared_configs.configs import INTENT_MODEL_TAG
-from shared_configs.configs import INTENT_MODEL_VERSION
-from shared_configs.model_server_models import ConnectorClassificationRequest
-from shared_configs.model_server_models import ConnectorClassificationResponse
-from shared_configs.model_server_models import ContentClassificationPrediction
-from shared_configs.model_server_models import IntentRequest
-from shared_configs.model_server_models import IntentResponse
-
-if TYPE_CHECKING:
-    from setfit import SetFitModel  # type: ignore[import-untyped]
-    from transformers import PreTrainedTokenizer, BatchEncoding
-
-
-logger = setup_logger()
-
-router = APIRouter(prefix="/custom")
-
-_CONNECTOR_CLASSIFIER_TOKENIZER: Optional["PreTrainedTokenizer"] = None
-_CONNECTOR_CLASSIFIER_MODEL: ConnectorClassifier | None = None
-
-_INTENT_TOKENIZER: Optional["PreTrainedTokenizer"] = None
-_INTENT_MODEL: HybridClassifier | None = None
-
-_INFORMATION_CONTENT_MODEL: Optional["SetFitModel"] = None
-
-_INFORMATION_CONTENT_MODEL_PROMPT_PREFIX: str = ""  # spec to model version!
-
-
-def get_connector_classifier_tokenizer() -> "PreTrainedTokenizer":
-    global _CONNECTOR_CLASSIFIER_TOKENIZER
-    from transformers import AutoTokenizer, PreTrainedTokenizer
-
-    if _CONNECTOR_CLASSIFIER_TOKENIZER is None:
-        # The tokenizer details are not uploaded to the HF hub since it's just the
-        # unmodified distilbert tokenizer.
-        _CONNECTOR_CLASSIFIER_TOKENIZER = cast(
-            PreTrainedTokenizer,
-            AutoTokenizer.from_pretrained("distilbert-base-uncased"),
-        )
-    return _CONNECTOR_CLASSIFIER_TOKENIZER
-
-
-def get_local_connector_classifier(
-    model_name_or_path: str = CONNECTOR_CLASSIFIER_MODEL_REPO,
-    tag: str = CONNECTOR_CLASSIFIER_MODEL_TAG,
-) -> ConnectorClassifier:
-    global _CONNECTOR_CLASSIFIER_MODEL
-    if _CONNECTOR_CLASSIFIER_MODEL is None:
-        try:
-            # Calculate where the cache should be, then load from local if available
-            local_path = snapshot_download(
-                repo_id=model_name_or_path, revision=tag, local_files_only=True
-            )
-            _CONNECTOR_CLASSIFIER_MODEL = ConnectorClassifier.from_pretrained(
-                local_path
-            )
-        except Exception as e:
-            logger.warning(f"Failed to load model directly: {e}")
-            try:
-                # Attempt to download the model snapshot
-                logger.info(f"Downloading model snapshot for {model_name_or_path}")
-                local_path = snapshot_download(repo_id=model_name_or_path, revision=tag)
-                _CONNECTOR_CLASSIFIER_MODEL = ConnectorClassifier.from_pretrained(
-                    local_path
-                )
-            except Exception as e:
-                logger.error(
-                    f"Failed to load model even after attempted snapshot download: {e}"
-                )
-                raise
-    return _CONNECTOR_CLASSIFIER_MODEL
-
-
-def get_intent_model_tokenizer() -> "PreTrainedTokenizer":
-    from transformers import AutoTokenizer, PreTrainedTokenizer
-
-    global _INTENT_TOKENIZER
-    if _INTENT_TOKENIZER is None:
-        # The tokenizer details are not uploaded to the HF hub since it's just the
-        # unmodified distilbert tokenizer.
-        _INTENT_TOKENIZER = cast(
-            PreTrainedTokenizer,
-            AutoTokenizer.from_pretrained("distilbert-base-uncased"),
-        )
-    return _INTENT_TOKENIZER
-
-
-def get_local_intent_model(
-    model_name_or_path: str = INTENT_MODEL_VERSION,
-    tag: str | None = INTENT_MODEL_TAG,
-) -> HybridClassifier:
-    global _INTENT_MODEL
-    if _INTENT_MODEL is None:
-        try:
-            # Calculate where the cache should be, then load from local if available
-            logger.notice(f"Loading model from local cache: {model_name_or_path}")
-            local_path = snapshot_download(
-                repo_id=model_name_or_path, revision=tag, local_files_only=True
-            )
-            _INTENT_MODEL = HybridClassifier.from_pretrained(local_path)
-            logger.notice(f"Loaded model from local cache: {local_path}")
-        except Exception as e:
-            logger.warning(f"Failed to load model directly: {e}")
-            try:
-                # Attempt to download the model snapshot
-                logger.notice(f"Downloading model snapshot for {model_name_or_path}")
-                local_path = snapshot_download(
-                    repo_id=model_name_or_path, revision=tag, local_files_only=False
-                )
-                _INTENT_MODEL = HybridClassifier.from_pretrained(local_path)
-            except Exception as e:
-                logger.error(
-                    f"Failed to load model even after attempted snapshot download: {e}"
-                )
-                raise
-    return _INTENT_MODEL
-
-
-def get_local_information_content_model(
-    model_name_or_path: str = INFORMATION_CONTENT_MODEL_VERSION,
-    tag: str | None = INFORMATION_CONTENT_MODEL_TAG,
-) -> "SetFitModel":
-    from setfit import SetFitModel
-
-    global _INFORMATION_CONTENT_MODEL
-    if _INFORMATION_CONTENT_MODEL is None:
-        try:
-            # Calculate where the cache should be, then load from local if available
-            logger.notice(
-                f"Loading content information model from local cache: {model_name_or_path}"
-            )
-            local_path = snapshot_download(
-                repo_id=model_name_or_path, revision=tag, local_files_only=True
-            )
-            _INFORMATION_CONTENT_MODEL = SetFitModel.from_pretrained(local_path)
-            logger.notice(
-                f"Loaded content information model from local cache: {local_path}"
-            )
-        except Exception as e:
-            logger.warning(f"Failed to load content information model directly: {e}")
-            try:
-                # Attempt to download the model snapshot
-                logger.notice(
-                    f"Downloading content information model snapshot for {model_name_or_path}"
-                )
-                local_path = snapshot_download(
-                    repo_id=model_name_or_path, revision=tag, local_files_only=False
-                )
-                _INFORMATION_CONTENT_MODEL = SetFitModel.from_pretrained(local_path)
-            except Exception as e:
-                logger.error(
-                    f"Failed to load content information model even after attempted snapshot download: {e}"
-                )
-                raise
-
-    return _INFORMATION_CONTENT_MODEL
-
-
-def tokenize_connector_classification_query(
-    connectors: list[str],
-    query: str,
-    tokenizer: "PreTrainedTokenizer",
-    connector_token_end_id: int,
-) -> tuple[torch.Tensor, torch.Tensor]:
-    """
-    Tokenize the connectors & user query into one prompt for the forward pass of ConnectorClassifier models
-
-    The attention mask is just all 1s. The prompt is CLS + each connector name suffixed with the connector end
-    token and then the user query.
-    """
-
-    input_ids = torch.tensor([tokenizer.cls_token_id], dtype=torch.long)
-
-    for connector in connectors:
-        connector_token_ids = tokenizer(
-            connector,
-            add_special_tokens=False,
-            return_tensors="pt",
-        )
-
-        input_ids = torch.cat(
-            (
-                input_ids,
-                connector_token_ids["input_ids"].squeeze(dim=0),
-                torch.tensor([connector_token_end_id], dtype=torch.long),
-            ),
-            dim=-1,
-        )
-    query_token_ids = tokenizer(
-        query,
-        add_special_tokens=False,
-        return_tensors="pt",
-    )
-
-    input_ids = torch.cat(
-        (
-            input_ids,
-            query_token_ids["input_ids"].squeeze(dim=0),
-            torch.tensor([tokenizer.sep_token_id], dtype=torch.long),
-        ),
-        dim=-1,
-    )
-    attention_mask = torch.ones(input_ids.numel(), dtype=torch.long)
-
-    return input_ids.unsqueeze(0), attention_mask.unsqueeze(0)
-
-
-def warm_up_connector_classifier_model() -> None:
-    logger.info(
-        f"Warming up connector_classifier model {CONNECTOR_CLASSIFIER_MODEL_TAG}"
-    )
-    connector_classifier_tokenizer = get_connector_classifier_tokenizer()
-    connector_classifier = get_local_connector_classifier()
-
-    input_ids, attention_mask = tokenize_connector_classification_query(
-        ["GitHub"],
-        "onyx classifier query google doc",
-        connector_classifier_tokenizer,
-        connector_classifier.connector_end_token_id,
-    )
-    input_ids = input_ids.to(connector_classifier.device)
-    attention_mask = attention_mask.to(connector_classifier.device)
-
-    connector_classifier(input_ids, attention_mask)
-
-
-def warm_up_intent_model() -> None:
-    logger.notice(f"Warming up Intent Model: {INTENT_MODEL_VERSION}")
-    intent_tokenizer = get_intent_model_tokenizer()
-    tokens = intent_tokenizer(
-        MODEL_WARM_UP_STRING, return_tensors="pt", truncation=True, padding=True
-    )
-
-    intent_model = get_local_intent_model()
-    device = intent_model.device
-    intent_model(
-        query_ids=tokens["input_ids"].to(device),
-        query_mask=tokens["attention_mask"].to(device),
-    )
-
-
-def warm_up_information_content_model() -> None:
-    logger.notice("Warming up Content Model")  # TODO: add version if needed
-
-    information_content_model = get_local_information_content_model()
-    information_content_model(INFORMATION_CONTENT_MODEL_WARM_UP_STRING)
-
-
-@simple_log_function_time()
-def run_inference(tokens: "BatchEncoding") -> tuple[list[float], list[float]]:
-    intent_model = get_local_intent_model()
-    device = intent_model.device
-
-    outputs = intent_model(
-        query_ids=tokens["input_ids"].to(device),
-        query_mask=tokens["attention_mask"].to(device),
-    )
-
-    token_logits = outputs["token_logits"]
-    intent_logits = outputs["intent_logits"]
-
-    # Move tensors to CPU before applying softmax and converting to numpy
-    intent_probabilities = F.softmax(intent_logits.cpu(), dim=-1).numpy()[0]
-    token_probabilities = F.softmax(token_logits.cpu(), dim=-1).numpy()[0]
-
-    # Extract the probabilities for the positive class (index 1) for each token
-    token_positive_probs = token_probabilities[:, 1].tolist()
-
-    return intent_probabilities.tolist(), token_positive_probs
-
-
-@simple_log_function_time()
-def run_content_classification_inference(
-    text_inputs: list[str],
-) -> list[ContentClassificationPrediction]:
-    """
-    Assign a score to the segments in question. The model stored in get_local_information_content_model()
-    creates the 'model score' based on its training, and the scores are then converted to a 0.0-1.0 scale.
-    In the code outside of the model/inference model servers that score will be converted into the actual
-    boost factor.
-    """
-
-    def _prob_to_score(prob: float) -> float:
-        """
-        Conversion of base score to 0.0 - 1.0 score. Note that the min/max values depend on the model!
-        """
-        _MIN_BASE_SCORE = 0.25
-        _MAX_BASE_SCORE = 0.75
-        if prob < _MIN_BASE_SCORE:
-            raw_score = 0.0
-        elif prob < _MAX_BASE_SCORE:
-            raw_score = (prob - _MIN_BASE_SCORE) / (_MAX_BASE_SCORE - _MIN_BASE_SCORE)
-        else:
-            raw_score = 1.0
-        return (
-            INDEXING_INFORMATION_CONTENT_CLASSIFICATION_MIN
-            + (
-                INDEXING_INFORMATION_CONTENT_CLASSIFICATION_MAX
-                - INDEXING_INFORMATION_CONTENT_CLASSIFICATION_MIN
-            )
-            * raw_score
-        )
-
-    _BATCH_SIZE = 32
-    content_model = get_local_information_content_model()
-
-    # Process inputs in batches
-    all_output_classes: list[int] = []
-    all_base_output_probabilities: list[float] = []
-
-    for i in range(0, len(text_inputs), _BATCH_SIZE):
-        batch = text_inputs[i : i + _BATCH_SIZE]
-        batch_with_prefix = []
-        batch_indices = []
-
-        # Pre-allocate results for this batch
-        batch_output_classes: list[np.ndarray] = [np.array(1)] * len(batch)
-        batch_probabilities: list[np.ndarray] = [np.array(1.0)] * len(batch)
-
-        # Pre-process batch to handle long input exceptions
-        for j, text in enumerate(batch):
-            if len(text) == 0:
-                # if no input, treat as non-informative from the model's perspective
-                batch_output_classes[j] = np.array(0)
-                batch_probabilities[j] = np.array(0.0)
-                logger.warning("Input for Content Information Model is empty")
-
-            elif (
-                len(text.split())
-                <= INDEXING_INFORMATION_CONTENT_CLASSIFICATION_CUTOFF_LENGTH
-            ):
-                # if input is short, use the model
-                batch_with_prefix.append(
-                    _INFORMATION_CONTENT_MODEL_PROMPT_PREFIX + text
-                )
-                batch_indices.append(j)
-            else:
-                # if longer than cutoff, treat as informative (stay with default), but issue warning
-                logger.warning("Input for Content Information Model too long")
-
-        if batch_with_prefix:  # Only run model if we have valid inputs
-            # Get predictions for the batch
-            model_output_classes = content_model(batch_with_prefix)
-            model_output_probabilities = content_model.predict_proba(batch_with_prefix)
-
-            # Place results in the correct positions
-            for idx, batch_idx in enumerate(batch_indices):
-                batch_output_classes[batch_idx] = model_output_classes[idx].numpy()
-                batch_probabilities[batch_idx] = model_output_probabilities[idx][
-                    1
-                ].numpy()  # x[1] is prob of the positive class
-
-        all_output_classes.extend([int(x) for x in batch_output_classes])
-        all_base_output_probabilities.extend([float(x) for x in batch_probabilities])
-
-    logits = [
-        np.log(p / (1 - p)) if p != 0.0 and p != 1.0 else (100 if p == 1.0 else -100)
-        for p in all_base_output_probabilities
-    ]
-    scaled_logits = [
-        logit / INDEXING_INFORMATION_CONTENT_CLASSIFICATION_TEMPERATURE
-        for logit in logits
-    ]
-    output_probabilities_with_temp = [
-        np.exp(scaled_logit) / (1 + np.exp(scaled_logit))
-        for scaled_logit in scaled_logits
-    ]
-
-    prediction_scores = [
-        _prob_to_score(p_temp) for p_temp in output_probabilities_with_temp
-    ]
-
-    content_classification_predictions = [
-        ContentClassificationPrediction(
-            predicted_label=predicted_label, content_boost_factor=output_score
-        )
-        for predicted_label, output_score in zip(all_output_classes, prediction_scores)
-    ]
-
-    return content_classification_predictions
-
-
-def map_keywords(
-    input_ids: torch.Tensor, tokenizer: "PreTrainedTokenizer", is_keyword: list[bool]
-) -> list[str]:
-    tokens = tokenizer.convert_ids_to_tokens(input_ids)  # type: ignore
-
-    if not len(tokens) == len(is_keyword):
-        raise ValueError("Length of tokens and keyword predictions must match")
-
-    if input_ids[0] == tokenizer.cls_token_id:
-        tokens = tokens[1:]
-        is_keyword = is_keyword[1:]
-
-    if input_ids[-1] == tokenizer.sep_token_id:
-        tokens = tokens[:-1]
-        is_keyword = is_keyword[:-1]
-
-    unk_token = tokenizer.unk_token
-    if unk_token in tokens:
-        raise ValueError("Unknown token detected in the input")
-
-    keywords = []
-    current_keyword = ""
-
-    for ind, token in enumerate(tokens):
-        if is_keyword[ind]:
-            if token.startswith("##"):
-                current_keyword += token[2:]
-            else:
-                if current_keyword:
-                    keywords.append(current_keyword)
-                current_keyword = token
-        else:
-            # If mispredicted a later token of a keyword, add it to the current keyword
-            # to complete it
-            if current_keyword:
-                if len(current_keyword) > 2 and current_keyword.startswith("##"):
-                    current_keyword = current_keyword[2:]
-
-                else:
-                    keywords.append(current_keyword)
-                    current_keyword = ""
-
-    if current_keyword:
-        keywords.append(current_keyword)
-
-    return keywords
-
-
-def clean_keywords(keywords: list[str]) -> list[str]:
-    cleaned_words = []
-    for word in keywords:
-        word = word[:-2] if word.endswith("'s") else word
-        word = word.replace("/", " ")
-        word = word.replace("'", "").replace('"', "")
-        cleaned_words.extend([w for w in word.strip().split() if w and not w.isspace()])
-    return cleaned_words
-
-
-def run_connector_classification(req: ConnectorClassificationRequest) -> list[str]:
-    tokenizer = get_connector_classifier_tokenizer()
-    model = get_local_connector_classifier()
-
-    connector_names = req.available_connectors
-
-    input_ids, attention_mask = tokenize_connector_classification_query(
-        connector_names,
-        req.query,
-        tokenizer,
-        model.connector_end_token_id,
-    )
-    input_ids = input_ids.to(model.device)
-    attention_mask = attention_mask.to(model.device)
-
-    global_confidence, classifier_confidence = model(input_ids, attention_mask)
-
-    if global_confidence.item() < 0.5:
-        return []
-
-    passed_connectors = []
-
-    for i, connector_name in enumerate(connector_names):
-        if classifier_confidence.view(-1)[i].item() > 0.5:
-            passed_connectors.append(connector_name)
-
-    return passed_connectors
-
-
-def run_analysis(intent_req: IntentRequest) -> tuple[bool, list[str]]:
-    tokenizer = get_intent_model_tokenizer()
-    model_input = tokenizer(
-        intent_req.query, return_tensors="pt", truncation=False, padding=False
-    )
-
-    if len(model_input.input_ids[0]) > 512:
-        # If the user text is too long, assume it is semantic and keep all words
-        return True, intent_req.query.split()
-
-    intent_probs, token_probs = run_inference(model_input)
-
-    is_keyword_sequence = intent_probs[0] >= intent_req.keyword_percent_threshold
-
-    keyword_preds = [
-        token_prob >= intent_req.keyword_percent_threshold for token_prob in token_probs
-    ]
-
-    try:
-        keywords = map_keywords(model_input.input_ids[0], tokenizer, keyword_preds)
-    except Exception as e:
-        logger.warning(
-            f"Failed to extract keywords for query: {intent_req.query} due to {e}"
-        )
-        # Fallback to keeping all words
-        keywords = intent_req.query.split()
-
-    cleaned_keywords = clean_keywords(keywords)
-
-    return is_keyword_sequence, cleaned_keywords
-
-
-@router.post("/connector-classification")
-async def process_connector_classification_request(
-    classification_request: ConnectorClassificationRequest,
-) -> ConnectorClassificationResponse:
-    if INDEXING_ONLY:
-        raise RuntimeError(
-            "Indexing model server should not call connector classification endpoint"
-        )
-
-    if len(classification_request.available_connectors) == 0:
-        return ConnectorClassificationResponse(connectors=[])
-
-    connectors = run_connector_classification(classification_request)
-    return ConnectorClassificationResponse(connectors=connectors)
-
-
-@router.post("/query-analysis")
-async def process_analysis_request(
-    intent_request: IntentRequest,
-) -> IntentResponse:
-    if INDEXING_ONLY:
-        raise RuntimeError("Indexing model server should not call intent endpoint")
-
-    is_keyword, keywords = run_analysis(intent_request)
-    return IntentResponse(is_keyword=is_keyword, keywords=keywords)
-
-
-@router.post("/content-classification")
-async def process_content_classification_request(
-    content_classification_requests: list[str],
-) -> list[ContentClassificationPrediction]:
-    return run_content_classification_inference(content_classification_requests)

backend/model_server/encoders.py

@@ -1,7 +1,6 @@
 import asyncio
 import time
 from typing import Any
-from typing import Optional
 from typing import TYPE_CHECKING
 
 from fastapi import APIRouter
@@ -10,16 +9,13 @@ from fastapi import Request
 
 from model_server.utils import simple_log_function_time
 from onyx.utils.logger import setup_logger
-from shared_configs.configs import INDEXING_ONLY
 from shared_configs.enums import EmbedTextType
 from shared_configs.model_server_models import Embedding
 from shared_configs.model_server_models import EmbedRequest
 from shared_configs.model_server_models import EmbedResponse
-from shared_configs.model_server_models import RerankRequest
-from shared_configs.model_server_models import RerankResponse
 
 if TYPE_CHECKING:
-    from sentence_transformers import CrossEncoder, SentenceTransformer
+    from sentence_transformers import SentenceTransformer
 
 logger = setup_logger()
 
@@ -27,11 +23,6 @@ router = APIRouter(prefix="/encoder")
 
 
 _GLOBAL_MODELS_DICT: dict[str, "SentenceTransformer"] = {}
-_RERANK_MODEL: Optional["CrossEncoder"] = None
-
-# If we are not only indexing, dont want retry very long
-_RETRY_DELAY = 10 if INDEXING_ONLY else 0.1
-_RETRY_TRIES = 10 if INDEXING_ONLY else 2
 
 
 def get_embedding_model(
@@ -87,19 +78,6 @@ def get_embedding_model(
     return _GLOBAL_MODELS_DICT[model_name]
 
 
-def get_local_reranking_model(
-    model_name: str,
-) -> "CrossEncoder":
-    global _RERANK_MODEL
-    from sentence_transformers import CrossEncoder
-
-    if _RERANK_MODEL is None:
-        logger.notice(f"Loading {model_name}")
-        model = CrossEncoder(model_name)
-        _RERANK_MODEL = model
-    return _RERANK_MODEL
-
-
 ENCODING_RETRIES = 3
 ENCODING_RETRY_DELAY = 0.1
 
@@ -189,16 +167,6 @@ async def embed_text(
     return embeddings
 
 
-@simple_log_function_time()
-async def local_rerank(query: str, docs: list[str], model_name: str) -> list[float]:
-    cross_encoder = get_local_reranking_model(model_name)
-    # Run CPU-bound reranking in a thread pool
-    return await asyncio.get_event_loop().run_in_executor(
-        None,
-        lambda: cross_encoder.predict([(query, doc) for doc in docs]).tolist(),
-    )
-
-
 @router.post("/bi-encoder-embed")
 async def route_bi_encoder_embed(
     request: Request,
@@ -254,39 +222,3 @@ async def process_embed_request(
         raise HTTPException(
             status_code=500, detail=f"Error during embedding process: {e}"
         )
-
-
-@router.post("/cross-encoder-scores")
-async def process_rerank_request(rerank_request: RerankRequest) -> RerankResponse:
-    """Cross encoders can be purely black box from the app perspective"""
-    # Only local models should use this endpoint - API providers should make direct API calls
-    if rerank_request.provider_type is not None:
-        raise ValueError(
-            f"Model server reranking endpoint should only be used for local models. "
-            f"API provider '{rerank_request.provider_type}' should make direct API calls instead."
-        )
-
-    if INDEXING_ONLY:
-        raise RuntimeError("Indexing model server should not call intent endpoint")
-
-    if not rerank_request.documents or not rerank_request.query:
-        raise HTTPException(
-            status_code=400, detail="Missing documents or query for reranking"
-        )
-    if not all(rerank_request.documents):
-        raise ValueError("Empty documents cannot be reranked.")
-
-    try:
-        # At this point, provider_type is None, so handle local reranking
-        sim_scores = await local_rerank(
-            query=rerank_request.query,
-            docs=rerank_request.documents,
-            model_name=rerank_request.model_name,
-        )
-        return RerankResponse(scores=sim_scores)
-
-    except Exception as e:
-        logger.exception(f"Error during reranking process:\n{str(e)}")
-        raise HTTPException(
-            status_code=500, detail="Failed to run Cross-Encoder reranking"
-        )

backend/model_server/legacy/README.md

@@ -0,0 +1,5 @@
+This directory contains code that was useful and may become useful again in the future.
+
+We stopped using rerankers because the state of the art rerankers are not significantly better than the biencoders and much worse than LLMs which are also capable of acting on a small set of documents for filtering, reranking, etc.
+
+We stopped using the internal query classifier as that's now offloaded to the LLM which does query expansion so we know ahead of time if it's a keyword or semantic query.

backend/model_server/legacy/custom_models.py

@@ -0,0 +1,573 @@
+# from typing import cast
+# from typing import Optional
+# from typing import TYPE_CHECKING
+
+# import numpy as np
+# import torch
+# import torch.nn.functional as F
+# from fastapi import APIRouter
+# from huggingface_hub import snapshot_download
+# from pydantic import BaseModel
+
+# from model_server.constants import MODEL_WARM_UP_STRING
+# from model_server.legacy.onyx_torch_model import ConnectorClassifier
+# from model_server.legacy.onyx_torch_model import HybridClassifier
+# from model_server.utils import simple_log_function_time
+# from onyx.utils.logger import setup_logger
+# from shared_configs.configs import CONNECTOR_CLASSIFIER_MODEL_REPO
+# from shared_configs.configs import CONNECTOR_CLASSIFIER_MODEL_TAG
+# from shared_configs.configs import INDEXING_ONLY
+# from shared_configs.configs import INTENT_MODEL_TAG
+# from shared_configs.configs import INTENT_MODEL_VERSION
+# from shared_configs.model_server_models import IntentRequest
+# from shared_configs.model_server_models import IntentResponse
+
+# if TYPE_CHECKING:
+#     from setfit import SetFitModel  # type: ignore[import-untyped]
+#     from transformers import PreTrainedTokenizer, BatchEncoding
+
+
+# INFORMATION_CONTENT_MODEL_WARM_UP_STRING = "hi" * 50
+
+# INDEXING_INFORMATION_CONTENT_CLASSIFICATION_MAX = 1.0
+# INDEXING_INFORMATION_CONTENT_CLASSIFICATION_MIN = 0.7
+# INDEXING_INFORMATION_CONTENT_CLASSIFICATION_TEMPERATURE = 4.0
+# INDEXING_INFORMATION_CONTENT_CLASSIFICATION_CUTOFF_LENGTH = 10
+# INFORMATION_CONTENT_MODEL_VERSION = "onyx-dot-app/information-content-model"
+# INFORMATION_CONTENT_MODEL_TAG: str | None = None
+
+
+# class ConnectorClassificationRequest(BaseModel):
+#     available_connectors: list[str]
+#     query: str
+
+
+# class ConnectorClassificationResponse(BaseModel):
+#     connectors: list[str]
+
+
+# class ContentClassificationPrediction(BaseModel):
+#     predicted_label: int
+#     content_boost_factor: float
+
+
+# logger = setup_logger()
+
+# router = APIRouter(prefix="/custom")
+
+# _CONNECTOR_CLASSIFIER_TOKENIZER: Optional["PreTrainedTokenizer"] = None
+# _CONNECTOR_CLASSIFIER_MODEL: ConnectorClassifier | None = None
+
+# _INTENT_TOKENIZER: Optional["PreTrainedTokenizer"] = None
+# _INTENT_MODEL: HybridClassifier | None = None
+
+# _INFORMATION_CONTENT_MODEL: Optional["SetFitModel"] = None
+
+# _INFORMATION_CONTENT_MODEL_PROMPT_PREFIX: str = ""  # spec to model version!
+
+
+# def get_connector_classifier_tokenizer() -> "PreTrainedTokenizer":
+#     global _CONNECTOR_CLASSIFIER_TOKENIZER
+#     from transformers import AutoTokenizer, PreTrainedTokenizer
+
+#     if _CONNECTOR_CLASSIFIER_TOKENIZER is None:
+#         # The tokenizer details are not uploaded to the HF hub since it's just the
+#         # unmodified distilbert tokenizer.
+#         _CONNECTOR_CLASSIFIER_TOKENIZER = cast(
+#             PreTrainedTokenizer,
+#             AutoTokenizer.from_pretrained("distilbert-base-uncased"),
+#         )
+#     return _CONNECTOR_CLASSIFIER_TOKENIZER
+
+
+# def get_local_connector_classifier(
+#     model_name_or_path: str = CONNECTOR_CLASSIFIER_MODEL_REPO,
+#     tag: str = CONNECTOR_CLASSIFIER_MODEL_TAG,
+# ) -> ConnectorClassifier:
+#     global _CONNECTOR_CLASSIFIER_MODEL
+#     if _CONNECTOR_CLASSIFIER_MODEL is None:
+#         try:
+#             # Calculate where the cache should be, then load from local if available
+#             local_path = snapshot_download(
+#                 repo_id=model_name_or_path, revision=tag, local_files_only=True
+#             )
+#             _CONNECTOR_CLASSIFIER_MODEL = ConnectorClassifier.from_pretrained(
+#                 local_path
+#             )
+#         except Exception as e:
+#             logger.warning(f"Failed to load model directly: {e}")
+#             try:
+#                 # Attempt to download the model snapshot
+#                 logger.info(f"Downloading model snapshot for {model_name_or_path}")
+#                 local_path = snapshot_download(repo_id=model_name_or_path, revision=tag)
+#                 _CONNECTOR_CLASSIFIER_MODEL = ConnectorClassifier.from_pretrained(
+#                     local_path
+#                 )
+#             except Exception as e:
+#                 logger.error(
+#                     f"Failed to load model even after attempted snapshot download: {e}"
+#                 )
+#                 raise
+#     return _CONNECTOR_CLASSIFIER_MODEL
+
+
+# def get_intent_model_tokenizer() -> "PreTrainedTokenizer":
+#     from transformers import AutoTokenizer, PreTrainedTokenizer
+
+#     global _INTENT_TOKENIZER
+#     if _INTENT_TOKENIZER is None:
+#         # The tokenizer details are not uploaded to the HF hub since it's just the
+#         # unmodified distilbert tokenizer.
+#         _INTENT_TOKENIZER = cast(
+#             PreTrainedTokenizer,
+#             AutoTokenizer.from_pretrained("distilbert-base-uncased"),
+#         )
+#     return _INTENT_TOKENIZER
+
+
+# def get_local_intent_model(
+#     model_name_or_path: str = INTENT_MODEL_VERSION,
+#     tag: str | None = INTENT_MODEL_TAG,
+# ) -> HybridClassifier:
+#     global _INTENT_MODEL
+#     if _INTENT_MODEL is None:
+#         try:
+#             # Calculate where the cache should be, then load from local if available
+#             logger.notice(f"Loading model from local cache: {model_name_or_path}")
+#             local_path = snapshot_download(
+#                 repo_id=model_name_or_path, revision=tag, local_files_only=True
+#             )
+#             _INTENT_MODEL = HybridClassifier.from_pretrained(local_path)
+#             logger.notice(f"Loaded model from local cache: {local_path}")
+#         except Exception as e:
+#             logger.warning(f"Failed to load model directly: {e}")
+#             try:
+#                 # Attempt to download the model snapshot
+#                 logger.notice(f"Downloading model snapshot for {model_name_or_path}")
+#                 local_path = snapshot_download(
+#                     repo_id=model_name_or_path, revision=tag, local_files_only=False
+#                 )
+#                 _INTENT_MODEL = HybridClassifier.from_pretrained(local_path)
+#             except Exception as e:
+#                 logger.error(
+#                     f"Failed to load model even after attempted snapshot download: {e}"
+#                 )
+#                 raise
+#     return _INTENT_MODEL
+
+
+# def get_local_information_content_model(
+#     model_name_or_path: str = INFORMATION_CONTENT_MODEL_VERSION,
+#     tag: str | None = INFORMATION_CONTENT_MODEL_TAG,
+# ) -> "SetFitModel":
+#     from setfit import SetFitModel
+
+#     global _INFORMATION_CONTENT_MODEL
+#     if _INFORMATION_CONTENT_MODEL is None:
+#         try:
+#             # Calculate where the cache should be, then load from local if available
+#             logger.notice(
+#                 f"Loading content information model from local cache: {model_name_or_path}"
+#             )
+#             local_path = snapshot_download(
+#                 repo_id=model_name_or_path, revision=tag, local_files_only=True
+#             )
+#             _INFORMATION_CONTENT_MODEL = SetFitModel.from_pretrained(local_path)
+#             logger.notice(
+#                 f"Loaded content information model from local cache: {local_path}"
+#             )
+#         except Exception as e:
+#             logger.warning(f"Failed to load content information model directly: {e}")
+#             try:
+#                 # Attempt to download the model snapshot
+#                 logger.notice(
+#                     f"Downloading content information model snapshot for {model_name_or_path}"
+#                 )
+#                 local_path = snapshot_download(
+#                     repo_id=model_name_or_path, revision=tag, local_files_only=False
+#                 )
+#                 _INFORMATION_CONTENT_MODEL = SetFitModel.from_pretrained(local_path)
+#             except Exception as e:
+#                 logger.error(
+#                     f"Failed to load content information model even after attempted snapshot download: {e}"
+#                 )
+#                 raise
+
+#     return _INFORMATION_CONTENT_MODEL
+
+
+# def tokenize_connector_classification_query(
+#     connectors: list[str],
+#     query: str,
+#     tokenizer: "PreTrainedTokenizer",
+#     connector_token_end_id: int,
+# ) -> tuple[torch.Tensor, torch.Tensor]:
+#     """
+#     Tokenize the connectors & user query into one prompt for the forward pass of ConnectorClassifier models
+
+#     The attention mask is just all 1s. The prompt is CLS + each connector name suffixed with the connector end
+#     token and then the user query.
+#     """
+
+#     input_ids = torch.tensor([tokenizer.cls_token_id], dtype=torch.long)
+
+#     for connector in connectors:
+#         connector_token_ids = tokenizer(
+#             connector,
+#             add_special_tokens=False,
+#             return_tensors="pt",
+#         )
+
+#         input_ids = torch.cat(
+#             (
+#                 input_ids,
+#                 connector_token_ids["input_ids"].squeeze(dim=0),
+#                 torch.tensor([connector_token_end_id], dtype=torch.long),
+#             ),
+#             dim=-1,
+#         )
+#     query_token_ids = tokenizer(
+#         query,
+#         add_special_tokens=False,
+#         return_tensors="pt",
+#     )
+
+#     input_ids = torch.cat(
+#         (
+#             input_ids,
+#             query_token_ids["input_ids"].squeeze(dim=0),
+#             torch.tensor([tokenizer.sep_token_id], dtype=torch.long),
+#         ),
+#         dim=-1,
+#     )
+#     attention_mask = torch.ones(input_ids.numel(), dtype=torch.long)
+
+#     return input_ids.unsqueeze(0), attention_mask.unsqueeze(0)
+
+
+# def warm_up_connector_classifier_model() -> None:
+#     logger.info(
+#         f"Warming up connector_classifier model {CONNECTOR_CLASSIFIER_MODEL_TAG}"
+#     )
+#     connector_classifier_tokenizer = get_connector_classifier_tokenizer()
+#     connector_classifier = get_local_connector_classifier()
+
+#     input_ids, attention_mask = tokenize_connector_classification_query(
+#         ["GitHub"],
+#         "onyx classifier query google doc",
+#         connector_classifier_tokenizer,
+#         connector_classifier.connector_end_token_id,
+#     )
+#     input_ids = input_ids.to(connector_classifier.device)
+#     attention_mask = attention_mask.to(connector_classifier.device)
+
+#     connector_classifier(input_ids, attention_mask)
+
+
+# def warm_up_intent_model() -> None:
+#     logger.notice(f"Warming up Intent Model: {INTENT_MODEL_VERSION}")
+#     intent_tokenizer = get_intent_model_tokenizer()
+#     tokens = intent_tokenizer(
+#         MODEL_WARM_UP_STRING, return_tensors="pt", truncation=True, padding=True
+#     )
+
+#     intent_model = get_local_intent_model()
+#     device = intent_model.device
+#     intent_model(
+#         query_ids=tokens["input_ids"].to(device),
+#         query_mask=tokens["attention_mask"].to(device),
+#     )
+
+
+# def warm_up_information_content_model() -> None:
+#     logger.notice("Warming up Content Model")  # TODO: add version if needed
+
+#     information_content_model = get_local_information_content_model()
+#     information_content_model(INFORMATION_CONTENT_MODEL_WARM_UP_STRING)
+
+
+# @simple_log_function_time()
+# def run_inference(tokens: "BatchEncoding") -> tuple[list[float], list[float]]:
+#     intent_model = get_local_intent_model()
+#     device = intent_model.device
+
+#     outputs = intent_model(
+#         query_ids=tokens["input_ids"].to(device),
+#         query_mask=tokens["attention_mask"].to(device),
+#     )
+
+#     token_logits = outputs["token_logits"]
+#     intent_logits = outputs["intent_logits"]
+
+#     # Move tensors to CPU before applying softmax and converting to numpy
+#     intent_probabilities = F.softmax(intent_logits.cpu(), dim=-1).numpy()[0]
+#     token_probabilities = F.softmax(token_logits.cpu(), dim=-1).numpy()[0]
+
+#     # Extract the probabilities for the positive class (index 1) for each token
+#     token_positive_probs = token_probabilities[:, 1].tolist()
+
+#     return intent_probabilities.tolist(), token_positive_probs
+
+
+# @simple_log_function_time()
+# def run_content_classification_inference(
+#     text_inputs: list[str],
+# ) -> list[ContentClassificationPrediction]:
+#     """
+#     Assign a score to the segments in question. The model stored in get_local_information_content_model()
+#     creates the 'model score' based on its training, and the scores are then converted to a 0.0-1.0 scale.
+#     In the code outside of the model/inference model servers that score will be converted into the actual
+#     boost factor.
+#     """
+
+#     def _prob_to_score(prob: float) -> float:
+#         """
+#         Conversion of base score to 0.0 - 1.0 score. Note that the min/max values depend on the model!
+#         """
+#         _MIN_BASE_SCORE = 0.25
+#         _MAX_BASE_SCORE = 0.75
+#         if prob < _MIN_BASE_SCORE:
+#             raw_score = 0.0
+#         elif prob < _MAX_BASE_SCORE:
+#             raw_score = (prob - _MIN_BASE_SCORE) / (_MAX_BASE_SCORE - _MIN_BASE_SCORE)
+#         else:
+#             raw_score = 1.0
+#         return (
+#             INDEXING_INFORMATION_CONTENT_CLASSIFICATION_MIN
+#             + (
+#                 INDEXING_INFORMATION_CONTENT_CLASSIFICATION_MAX
+#                 - INDEXING_INFORMATION_CONTENT_CLASSIFICATION_MIN
+#             )
+#             * raw_score
+#         )
+
+#     _BATCH_SIZE = 32
+#     content_model = get_local_information_content_model()
+
+#     # Process inputs in batches
+#     all_output_classes: list[int] = []
+#     all_base_output_probabilities: list[float] = []
+
+#     for i in range(0, len(text_inputs), _BATCH_SIZE):
+#         batch = text_inputs[i : i + _BATCH_SIZE]
+#         batch_with_prefix = []
+#         batch_indices = []
+
+#         # Pre-allocate results for this batch
+#         batch_output_classes: list[np.ndarray] = [np.array(1)] * len(batch)
+#         batch_probabilities: list[np.ndarray] = [np.array(1.0)] * len(batch)
+
+#         # Pre-process batch to handle long input exceptions
+#         for j, text in enumerate(batch):
+#             if len(text) == 0:
+#                 # if no input, treat as non-informative from the model's perspective
+#                 batch_output_classes[j] = np.array(0)
+#                 batch_probabilities[j] = np.array(0.0)
+#                 logger.warning("Input for Content Information Model is empty")
+
+#             elif (
+#                 len(text.split())
+#                 <= INDEXING_INFORMATION_CONTENT_CLASSIFICATION_CUTOFF_LENGTH
+#             ):
+#                 # if input is short, use the model
+#                 batch_with_prefix.append(
+#                     _INFORMATION_CONTENT_MODEL_PROMPT_PREFIX + text
+#                 )
+#                 batch_indices.append(j)
+#             else:
+#                 # if longer than cutoff, treat as informative (stay with default), but issue warning
+#                 logger.warning("Input for Content Information Model too long")
+
+#         if batch_with_prefix:  # Only run model if we have valid inputs
+#             # Get predictions for the batch
+#             model_output_classes = content_model(batch_with_prefix)
+#             model_output_probabilities = content_model.predict_proba(batch_with_prefix)
+
+#             # Place results in the correct positions
+#             for idx, batch_idx in enumerate(batch_indices):
+#                 batch_output_classes[batch_idx] = model_output_classes[idx].numpy()
+#                 batch_probabilities[batch_idx] = model_output_probabilities[idx][
+#                     1
+#                 ].numpy()  # x[1] is prob of the positive class
+
+#         all_output_classes.extend([int(x) for x in batch_output_classes])
+#         all_base_output_probabilities.extend([float(x) for x in batch_probabilities])
+
+#     logits = [
+#         np.log(p / (1 - p)) if p != 0.0 and p != 1.0 else (100 if p == 1.0 else -100)
+#         for p in all_base_output_probabilities
+#     ]
+#     scaled_logits = [
+#         logit / INDEXING_INFORMATION_CONTENT_CLASSIFICATION_TEMPERATURE
+#         for logit in logits
+#     ]
+#     output_probabilities_with_temp = [
+#         np.exp(scaled_logit) / (1 + np.exp(scaled_logit))
+#         for scaled_logit in scaled_logits
+#     ]
+
+#     prediction_scores = [
+#         _prob_to_score(p_temp) for p_temp in output_probabilities_with_temp
+#     ]
+
+#     content_classification_predictions = [
+#         ContentClassificationPrediction(
+#             predicted_label=predicted_label, content_boost_factor=output_score
+#         )
+#         for predicted_label, output_score in zip(all_output_classes, prediction_scores)
+#     ]
+
+#     return content_classification_predictions
+
+
+# def map_keywords(
+#     input_ids: torch.Tensor, tokenizer: "PreTrainedTokenizer", is_keyword: list[bool]
+# ) -> list[str]:
+#     tokens = tokenizer.convert_ids_to_tokens(input_ids)  # type: ignore
+
+#     if not len(tokens) == len(is_keyword):
+#         raise ValueError("Length of tokens and keyword predictions must match")
+
+#     if input_ids[0] == tokenizer.cls_token_id:
+#         tokens = tokens[1:]
+#         is_keyword = is_keyword[1:]
+
+#     if input_ids[-1] == tokenizer.sep_token_id:
+#         tokens = tokens[:-1]
+#         is_keyword = is_keyword[:-1]
+
+#     unk_token = tokenizer.unk_token
+#     if unk_token in tokens:
+#         raise ValueError("Unknown token detected in the input")
+
+#     keywords = []
+#     current_keyword = ""
+
+#     for ind, token in enumerate(tokens):
+#         if is_keyword[ind]:
+#             if token.startswith("##"):
+#                 current_keyword += token[2:]
+#             else:
+#                 if current_keyword:
+#                     keywords.append(current_keyword)
+#                 current_keyword = token
+#         else:
+#             # If mispredicted a later token of a keyword, add it to the current keyword
+#             # to complete it
+#             if current_keyword:
+#                 if len(current_keyword) > 2 and current_keyword.startswith("##"):
+#                     current_keyword = current_keyword[2:]
+
+#                 else:
+#                     keywords.append(current_keyword)
+#                     current_keyword = ""
+
+#     if current_keyword:
+#         keywords.append(current_keyword)
+
+#     return keywords
+
+
+# def clean_keywords(keywords: list[str]) -> list[str]:
+#     cleaned_words = []
+#     for word in keywords:
+#         word = word[:-2] if word.endswith("'s") else word
+#         word = word.replace("/", " ")
+#         word = word.replace("'", "").replace('"', "")
+#         cleaned_words.extend([w for w in word.strip().split() if w and not w.isspace()])
+#     return cleaned_words
+
+
+# def run_connector_classification(req: ConnectorClassificationRequest) -> list[str]:
+#     tokenizer = get_connector_classifier_tokenizer()
+#     model = get_local_connector_classifier()
+
+#     connector_names = req.available_connectors
+
+#     input_ids, attention_mask = tokenize_connector_classification_query(
+#         connector_names,
+#         req.query,
+#         tokenizer,
+#         model.connector_end_token_id,
+#     )
+#     input_ids = input_ids.to(model.device)
+#     attention_mask = attention_mask.to(model.device)
+
+#     global_confidence, classifier_confidence = model(input_ids, attention_mask)
+
+#     if global_confidence.item() < 0.5:
+#         return []
+
+#     passed_connectors = []
+
+#     for i, connector_name in enumerate(connector_names):
+#         if classifier_confidence.view(-1)[i].item() > 0.5:
+#             passed_connectors.append(connector_name)
+
+#     return passed_connectors
+
+
+# def run_analysis(intent_req: IntentRequest) -> tuple[bool, list[str]]:
+#     tokenizer = get_intent_model_tokenizer()
+#     model_input = tokenizer(
+#         intent_req.query, return_tensors="pt", truncation=False, padding=False
+#     )
+
+#     if len(model_input.input_ids[0]) > 512:
+#         # If the user text is too long, assume it is semantic and keep all words
+#         return True, intent_req.query.split()
+
+#     intent_probs, token_probs = run_inference(model_input)
+
+#     is_keyword_sequence = intent_probs[0] >= intent_req.keyword_percent_threshold
+
+#     keyword_preds = [
+#         token_prob >= intent_req.keyword_percent_threshold for token_prob in token_probs
+#     ]
+
+#     try:
+#         keywords = map_keywords(model_input.input_ids[0], tokenizer, keyword_preds)
+#     except Exception as e:
+#         logger.warning(
+#             f"Failed to extract keywords for query: {intent_req.query} due to {e}"
+#         )
+#         # Fallback to keeping all words
+#         keywords = intent_req.query.split()
+
+#     cleaned_keywords = clean_keywords(keywords)
+
+#     return is_keyword_sequence, cleaned_keywords
+
+
+# @router.post("/connector-classification")
+# async def process_connector_classification_request(
+#     classification_request: ConnectorClassificationRequest,
+# ) -> ConnectorClassificationResponse:
+#     if INDEXING_ONLY:
+#         raise RuntimeError(
+#             "Indexing model server should not call connector classification endpoint"
+#         )
+
+#     if len(classification_request.available_connectors) == 0:
+#         return ConnectorClassificationResponse(connectors=[])
+
+#     connectors = run_connector_classification(classification_request)
+#     return ConnectorClassificationResponse(connectors=connectors)
+
+
+# @router.post("/query-analysis")
+# async def process_analysis_request(
+#     intent_request: IntentRequest,
+# ) -> IntentResponse:
+#     if INDEXING_ONLY:
+#         raise RuntimeError("Indexing model server should not call intent endpoint")
+
+#     is_keyword, keywords = run_analysis(intent_request)
+#     return IntentResponse(is_keyword=is_keyword, keywords=keywords)
+
+
+# @router.post("/content-classification")
+# async def process_content_classification_request(
+#     content_classification_requests: list[str],
+# ) -> list[ContentClassificationPrediction]:
+#     return run_content_classification_inference(content_classification_requests)

backend/model_server/legacy/onyx_torch_model.py

@@ -0,0 +1,154 @@
+# import json
+# import os
+# from typing import cast
+# from typing import TYPE_CHECKING
+
+# import torch
+# import torch.nn as nn
+
+
+# if TYPE_CHECKING:
+#     from transformers import DistilBertConfig
+
+
+# class HybridClassifier(nn.Module):
+#     def __init__(self) -> None:
+#         from transformers import DistilBertConfig, DistilBertModel
+
+#         super().__init__()
+#         config = DistilBertConfig()
+#         self.distilbert = DistilBertModel(config)
+#         config = self.distilbert.config  # type: ignore
+
+#         # Keyword tokenwise binary classification layer
+#         self.keyword_classifier = nn.Linear(config.dim, 2)
+
+#         # Intent Classifier layers
+#         self.pre_classifier = nn.Linear(config.dim, config.dim)
+#         self.intent_classifier = nn.Linear(config.dim, 2)
+
+#         self.device = torch.device("cpu")
+
+#     def forward(
+#         self,
+#         query_ids: torch.Tensor,
+#         query_mask: torch.Tensor,
+#     ) -> dict[str, torch.Tensor]:
+#         outputs = self.distilbert(input_ids=query_ids, attention_mask=query_mask)
+#         sequence_output = outputs.last_hidden_state
+
+#         # Intent classification on the CLS token
+#         cls_token_state = sequence_output[:, 0, :]
+#         pre_classifier_out = self.pre_classifier(cls_token_state)
+#         intent_logits = self.intent_classifier(pre_classifier_out)
+
+#         # Keyword classification on all tokens
+#         token_logits = self.keyword_classifier(sequence_output)
+
+#         return {"intent_logits": intent_logits, "token_logits": token_logits}
+
+#     @classmethod
+#     def from_pretrained(cls, load_directory: str) -> "HybridClassifier":
+#         model_path = os.path.join(load_directory, "pytorch_model.bin")
+#         config_path = os.path.join(load_directory, "config.json")
+
+#         with open(config_path, "r") as f:
+#             config = json.load(f)
+#         model = cls(**config)
+
+#         if torch.backends.mps.is_available():
+#             # Apple silicon GPU
+#             device = torch.device("mps")
+#         elif torch.cuda.is_available():
+#             device = torch.device("cuda")
+#         else:
+#             device = torch.device("cpu")
+
+#         model.load_state_dict(torch.load(model_path, map_location=device))
+#         model = model.to(device)
+
+#         model.device = device
+
+#         model.eval()
+#         # Eval doesn't set requires_grad to False, do it manually to save memory and have faster inference
+#         for param in model.parameters():
+#             param.requires_grad = False
+
+#         return model
+
+
+# class ConnectorClassifier(nn.Module):
+#     def __init__(self, config: "DistilBertConfig") -> None:
+#         from transformers import DistilBertTokenizer, DistilBertModel
+
+#         super().__init__()
+
+#         self.config = config
+#         self.distilbert = DistilBertModel(config)
+#         config = self.distilbert.config  # type: ignore
+#         self.connector_global_classifier = nn.Linear(config.dim, 1)
+#         self.connector_match_classifier = nn.Linear(config.dim, 1)
+#         self.tokenizer = DistilBertTokenizer.from_pretrained("distilbert-base-uncased")
+
+#         # Token indicating end of connector name, and on which classifier is used
+#         self.connector_end_token_id = self.tokenizer.get_vocab()[
+#             self.config.connector_end_token
+#         ]
+
+#         self.device = torch.device("cpu")
+
+#     def forward(
+#         self,
+#         input_ids: torch.Tensor,
+#         attention_mask: torch.Tensor,
+#     ) -> tuple[torch.Tensor, torch.Tensor]:
+#         hidden_states = self.distilbert(
+#             input_ids=input_ids, attention_mask=attention_mask
+#         ).last_hidden_state
+
+#         cls_hidden_states = hidden_states[
+#             :, 0, :
+#         ]  # Take leap of faith that first token is always [CLS]
+#         global_logits = self.connector_global_classifier(cls_hidden_states).view(-1)
+#         global_confidence = torch.sigmoid(global_logits).view(-1)
+
+#         connector_end_position_ids = input_ids == self.connector_end_token_id
+#         connector_end_hidden_states = hidden_states[connector_end_position_ids]
+#         classifier_output = self.connector_match_classifier(connector_end_hidden_states)
+#         classifier_confidence = torch.nn.functional.sigmoid(classifier_output).view(-1)
+
+#         return global_confidence, classifier_confidence
+
+#     @classmethod
+#     def from_pretrained(cls, repo_dir: str) -> "ConnectorClassifier":
+#         from transformers import DistilBertConfig
+
+#         config = cast(
+#             DistilBertConfig,
+#             DistilBertConfig.from_pretrained(os.path.join(repo_dir, "config.json")),
+#         )
+#         device = (
+#             torch.device("cuda")
+#             if torch.cuda.is_available()
+#             else (
+#                 torch.device("mps")
+#                 if torch.backends.mps.is_available()
+#                 else torch.device("cpu")
+#             )
+#         )
+#         state_dict = torch.load(
+#             os.path.join(repo_dir, "pytorch_model.pt"),
+#             map_location=device,
+#             weights_only=True,
+#         )
+
+#         model = cls(config)
+#         model.load_state_dict(state_dict)
+#         model.to(device)
+#         model.device = device
+#         model.eval()
+
+#         for param in model.parameters():
+#             param.requires_grad = False
+
+#         return model

backend/model_server/legacy/reranker.py

@@ -0,0 +1,80 @@
+# import asyncio
+# from typing import Optional
+# from typing import TYPE_CHECKING
+
+# from fastapi import APIRouter
+# from fastapi import HTTPException
+
+# from model_server.utils import simple_log_function_time
+# from onyx.utils.logger import setup_logger
+# from shared_configs.configs import INDEXING_ONLY
+# from shared_configs.model_server_models import RerankRequest
+# from shared_configs.model_server_models import RerankResponse
+
+# if TYPE_CHECKING:
+#     from sentence_transformers import CrossEncoder
+
+# logger = setup_logger()
+
+# router = APIRouter(prefix="/encoder")
+
+# _RERANK_MODEL: Optional["CrossEncoder"] = None
+
+
+# def get_local_reranking_model(
+#     model_name: str,
+# ) -> "CrossEncoder":
+#     global _RERANK_MODEL
+#     from sentence_transformers import CrossEncoder
+
+#     if _RERANK_MODEL is None:
+#         logger.notice(f"Loading {model_name}")
+#         model = CrossEncoder(model_name)
+#         _RERANK_MODEL = model
+#     return _RERANK_MODEL
+
+
+# @simple_log_function_time()
+# async def local_rerank(query: str, docs: list[str], model_name: str) -> list[float]:
+#     cross_encoder = get_local_reranking_model(model_name)
+#     # Run CPU-bound reranking in a thread pool
+#     return await asyncio.get_event_loop().run_in_executor(
+#         None,
+#         lambda: cross_encoder.predict([(query, doc) for doc in docs]).tolist(),
+#     )
+
+
+# @router.post("/cross-encoder-scores")
+# async def process_rerank_request(rerank_request: RerankRequest) -> RerankResponse:
+#     """Cross encoders can be purely black box from the app perspective"""
+#     # Only local models should use this endpoint - API providers should make direct API calls
+#     if rerank_request.provider_type is not None:
+#         raise ValueError(
+#             f"Model server reranking endpoint should only be used for local models. "
+#             f"API provider '{rerank_request.provider_type}' should make direct API calls instead."
+#         )
+
+#     if INDEXING_ONLY:
+#         raise RuntimeError("Indexing model server should not call reranking endpoint")
+
+#     if not rerank_request.documents or not rerank_request.query:
+#         raise HTTPException(
+#             status_code=400, detail="Missing documents or query for reranking"
+#         )
+#     if not all(rerank_request.documents):
+#         raise ValueError("Empty documents cannot be reranked.")
+
+#     try:
+#         # At this point, provider_type is None, so handle local reranking
+#         sim_scores = await local_rerank(
+#             query=rerank_request.query,
+#             docs=rerank_request.documents,
+#             model_name=rerank_request.model_name,
+#         )
+#         return RerankResponse(scores=sim_scores)
+
+#     except Exception as e:
+#         logger.exception(f"Error during reranking process:\n{str(e)}")
+#         raise HTTPException(
+#             status_code=500, detail="Failed to run Cross-Encoder reranking"
+#         )

backend/model_server/main.py

@@ -14,9 +14,6 @@ from sentry_sdk.integrations.fastapi import FastApiIntegration
 from sentry_sdk.integrations.starlette import StarletteIntegration
 from transformers import logging as transformer_logging
 
-from model_server.custom_models import router as custom_models_router
-from model_server.custom_models import warm_up_information_content_model
-from model_server.custom_models import warm_up_intent_model
 from model_server.encoders import router as encoders_router
 from model_server.management_endpoints import router as management_router
 from model_server.utils import get_gpu_type
@@ -30,7 +27,6 @@ from shared_configs.configs import MIN_THREADS_ML_MODELS
 from shared_configs.configs import MODEL_SERVER_ALLOWED_HOST
 from shared_configs.configs import MODEL_SERVER_PORT
 from shared_configs.configs import SENTRY_DSN
-from shared_configs.configs import SKIP_WARM_UP
 
 os.environ["TOKENIZERS_PARALLELISM"] = "false"
 os.environ["HF_HUB_DISABLE_TELEMETRY"] = "1"
@@ -92,18 +88,6 @@ async def lifespan(app: FastAPI) -> AsyncGenerator:
     torch.set_num_threads(max(MIN_THREADS_ML_MODELS, torch.get_num_threads()))
     logger.notice(f"Torch Threads: {torch.get_num_threads()}")
 
-    if not SKIP_WARM_UP:
-        if not INDEXING_ONLY:
-            logger.notice("Warming up intent model for inference model server")
-            warm_up_intent_model()
-        else:
-            logger.notice(
-                "Warming up content information model for indexing model server"
-            )
-            warm_up_information_content_model()
-    else:
-        logger.notice("Skipping model warmup due to SKIP_WARM_UP=true")
-
     yield
 
 
@@ -123,7 +107,6 @@ def get_model_app() -> FastAPI:
 
     application.include_router(management_router)
     application.include_router(encoders_router)
-    application.include_router(custom_models_router)
 
     request_id_prefix = "INF"
     if INDEXING_ONLY:

backend/model_server/onyx_torch_model.py

@@ -1,154 +0,0 @@
-import json
-import os
-from typing import cast
-from typing import TYPE_CHECKING
-
-import torch
-import torch.nn as nn
-
-
-if TYPE_CHECKING:
-    from transformers import DistilBertConfig
-
-
-class HybridClassifier(nn.Module):
-    def __init__(self) -> None:
-        from transformers import DistilBertConfig, DistilBertModel
-
-        super().__init__()
-        config = DistilBertConfig()
-        self.distilbert = DistilBertModel(config)
-        config = self.distilbert.config  # type: ignore
-
-        # Keyword tokenwise binary classification layer
-        self.keyword_classifier = nn.Linear(config.dim, 2)
-
-        # Intent Classifier layers
-        self.pre_classifier = nn.Linear(config.dim, config.dim)
-        self.intent_classifier = nn.Linear(config.dim, 2)
-
-        self.device = torch.device("cpu")
-
-    def forward(
-        self,
-        query_ids: torch.Tensor,
-        query_mask: torch.Tensor,
-    ) -> dict[str, torch.Tensor]:
-        outputs = self.distilbert(input_ids=query_ids, attention_mask=query_mask)
-        sequence_output = outputs.last_hidden_state
-
-        # Intent classification on the CLS token
-        cls_token_state = sequence_output[:, 0, :]
-        pre_classifier_out = self.pre_classifier(cls_token_state)
-        intent_logits = self.intent_classifier(pre_classifier_out)
-
-        # Keyword classification on all tokens
-        token_logits = self.keyword_classifier(sequence_output)
-
-        return {"intent_logits": intent_logits, "token_logits": token_logits}
-
-    @classmethod
-    def from_pretrained(cls, load_directory: str) -> "HybridClassifier":
-        model_path = os.path.join(load_directory, "pytorch_model.bin")
-        config_path = os.path.join(load_directory, "config.json")
-
-        with open(config_path, "r") as f:
-            config = json.load(f)
-        model = cls(**config)
-
-        if torch.backends.mps.is_available():
-            # Apple silicon GPU
-            device = torch.device("mps")
-        elif torch.cuda.is_available():
-            device = torch.device("cuda")
-        else:
-            device = torch.device("cpu")
-
-        model.load_state_dict(torch.load(model_path, map_location=device))
-        model = model.to(device)
-
-        model.device = device
-
-        model.eval()
-        # Eval doesn't set requires_grad to False, do it manually to save memory and have faster inference
-        for param in model.parameters():
-            param.requires_grad = False
-
-        return model
-
-
-class ConnectorClassifier(nn.Module):
-    def __init__(self, config: "DistilBertConfig") -> None:
-        from transformers import DistilBertTokenizer, DistilBertModel
-
-        super().__init__()
-
-        self.config = config
-        self.distilbert = DistilBertModel(config)
-        config = self.distilbert.config  # type: ignore
-        self.connector_global_classifier = nn.Linear(config.dim, 1)
-        self.connector_match_classifier = nn.Linear(config.dim, 1)
-        self.tokenizer = DistilBertTokenizer.from_pretrained("distilbert-base-uncased")
-
-        # Token indicating end of connector name, and on which classifier is used
-        self.connector_end_token_id = self.tokenizer.get_vocab()[
-            self.config.connector_end_token
-        ]
-
-        self.device = torch.device("cpu")
-
-    def forward(
-        self,
-        input_ids: torch.Tensor,
-        attention_mask: torch.Tensor,
-    ) -> tuple[torch.Tensor, torch.Tensor]:
-        hidden_states = self.distilbert(
-            input_ids=input_ids, attention_mask=attention_mask
-        ).last_hidden_state
-
-        cls_hidden_states = hidden_states[
-            :, 0, :
-        ]  # Take leap of faith that first token is always [CLS]
-        global_logits = self.connector_global_classifier(cls_hidden_states).view(-1)
-        global_confidence = torch.sigmoid(global_logits).view(-1)
-
-        connector_end_position_ids = input_ids == self.connector_end_token_id
-        connector_end_hidden_states = hidden_states[connector_end_position_ids]
-        classifier_output = self.connector_match_classifier(connector_end_hidden_states)
-        classifier_confidence = torch.nn.functional.sigmoid(classifier_output).view(-1)
-
-        return global_confidence, classifier_confidence
-
-    @classmethod
-    def from_pretrained(cls, repo_dir: str) -> "ConnectorClassifier":
-        from transformers import DistilBertConfig
-
-        config = cast(
-            DistilBertConfig,
-            DistilBertConfig.from_pretrained(os.path.join(repo_dir, "config.json")),
-        )
-        device = (
-            torch.device("cuda")
-            if torch.cuda.is_available()
-            else (
-                torch.device("mps")
-                if torch.backends.mps.is_available()
-                else torch.device("cpu")
-            )
-        )
-        state_dict = torch.load(
-            os.path.join(repo_dir, "pytorch_model.pt"),
-            map_location=device,
-            weights_only=True,
-        )
-
-        model = cls(config)
-        model.load_state_dict(state_dict)
-        model.to(device)
-        model.device = device
-        model.eval()
-
-        for param in model.parameters():
-            param.requires_grad = False
-
-        return model

backend/onyx/auth/captcha.py

@@ -0,0 +1,107 @@
+"""Captcha verification for user registration."""
+
+import httpx
+from pydantic import BaseModel
+from pydantic import Field
+
+from onyx.configs.app_configs import CAPTCHA_ENABLED
+from onyx.configs.app_configs import RECAPTCHA_SCORE_THRESHOLD
+from onyx.configs.app_configs import RECAPTCHA_SECRET_KEY
+from onyx.utils.logger import setup_logger
+
+logger = setup_logger()
+
+RECAPTCHA_VERIFY_URL = "https://www.google.com/recaptcha/api/siteverify"
+
+
+class CaptchaVerificationError(Exception):
+    """Raised when captcha verification fails."""
+
+
+class RecaptchaResponse(BaseModel):
+    """Response from Google reCAPTCHA verification API."""
+
+    success: bool
+    score: float | None = None  # Only present for reCAPTCHA v3
+    action: str | None = None
+    challenge_ts: str | None = None
+    hostname: str | None = None
+    error_codes: list[str] | None = Field(default=None, alias="error-codes")
+
+
+def is_captcha_enabled() -> bool:
+    """Check if captcha verification is enabled."""
+    return CAPTCHA_ENABLED and bool(RECAPTCHA_SECRET_KEY)
+
+
+async def verify_captcha_token(
+    token: str,
+    expected_action: str = "signup",
+) -> None:
+    """
+    Verify a reCAPTCHA token with Google's API.
+
+    Args:
+        token: The reCAPTCHA response token from the client
+        expected_action: Expected action name for v3 verification
+
+    Raises:
+        CaptchaVerificationError: If verification fails
+    """
+    if not is_captcha_enabled():
+        return
+
+    if not token:
+        raise CaptchaVerificationError("Captcha token is required")
+
+    try:
+        async with httpx.AsyncClient() as client:
+            response = await client.post(
+                RECAPTCHA_VERIFY_URL,
+                data={
+                    "secret": RECAPTCHA_SECRET_KEY,
+                    "response": token,
+                },
+                timeout=10.0,
+            )
+            response.raise_for_status()
+
+            data = response.json()
+            result = RecaptchaResponse(**data)
+
+            if not result.success:
+                error_codes = result.error_codes or ["unknown-error"]
+                logger.warning(f"Captcha verification failed: {error_codes}")
+                raise CaptchaVerificationError(
+                    f"Captcha verification failed: {', '.join(error_codes)}"
+                )
+
+            # For reCAPTCHA v3, also check the score
+            if result.score is not None:
+                if result.score < RECAPTCHA_SCORE_THRESHOLD:
+                    logger.warning(
+                        f"Captcha score too low: {result.score} < {RECAPTCHA_SCORE_THRESHOLD}"
+                    )
+                    raise CaptchaVerificationError(
+                        "Captcha verification failed: suspicious activity detected"
+                    )
+
+                # Optionally verify the action matches
+                if result.action and result.action != expected_action:
+                    logger.warning(
+                        f"Captcha action mismatch: {result.action} != {expected_action}"
+                    )
+                    raise CaptchaVerificationError(
+                        "Captcha verification failed: action mismatch"
+                    )
+
+            logger.debug(
+                f"Captcha verification passed: score={result.score}, "
+                f"action={result.action}"
+            )
+
+    except httpx.HTTPError as e:
+        logger.error(f"Captcha API request failed: {e}")
+        # In case of API errors, we might want to allow registration
+        # to prevent blocking legitimate users. This is a policy decision.
+        raise CaptchaVerificationError("Captcha verification service unavailable")

backend/onyx/auth/disposable_email_validator.py

@@ -0,0 +1,192 @@
+"""
+Utility to validate and block disposable/temporary email addresses.
+
+This module fetches a list of known disposable email domains from a remote source
+and caches them for performance. It's used during user registration to prevent
+abuse from temporary email services.
+"""
+
+import threading
+import time
+from typing import Set
+
+import httpx
+
+from onyx.configs.app_configs import DISPOSABLE_EMAIL_DOMAINS_URL
+from onyx.utils.logger import setup_logger
+
+logger = setup_logger()
+
+
+class DisposableEmailValidator:
+    """
+    Thread-safe singleton validator for disposable email domains.
+
+    Fetches and caches the list of disposable domains, with periodic refresh.
+    """
+
+    _instance: "DisposableEmailValidator | None" = None
+    _lock = threading.Lock()
+
+    def __new__(cls) -> "DisposableEmailValidator":
+        if cls._instance is None:
+            with cls._lock:
+                if cls._instance is None:
+                    cls._instance = super().__new__(cls)
+        return cls._instance
+
+    def __init__(self) -> None:
+        # Check if already initialized using a try/except to avoid type issues
+        try:
+            if self._initialized:
+                return
+        except AttributeError:
+            pass
+
+        self._domains: Set[str] = set()
+        self._last_fetch_time: float = 0
+        self._fetch_lock = threading.Lock()
+        # Cache for 1 hour
+        self._cache_duration = 3600
+        # Hardcoded fallback list of common disposable domains
+        # This ensures we block at least these even if the remote fetch fails
+        self._fallback_domains = {
+            "trashlify.com",
+            "10minutemail.com",
+            "guerrillamail.com",
+            "mailinator.com",
+            "tempmail.com",
+            "throwaway.email",
+            "yopmail.com",
+            "temp-mail.org",
+            "getnada.com",
+            "maildrop.cc",
+        }
+        # Set initialized flag last to prevent race conditions
+        self._initialized: bool = True
+
+    def _should_refresh(self) -> bool:
+        """Check if the cached domains should be refreshed."""
+        return (time.time() - self._last_fetch_time) > self._cache_duration
+
+    def _fetch_domains(self) -> Set[str]:
+        """
+        Fetch disposable email domains from the configured URL.
+
+        Returns:
+            Set of domain strings (lowercased)
+        """
+        if not DISPOSABLE_EMAIL_DOMAINS_URL:
+            logger.debug("DISPOSABLE_EMAIL_DOMAINS_URL not configured")
+            return self._fallback_domains.copy()
+
+        try:
+            logger.info(
+                f"Fetching disposable email domains from {DISPOSABLE_EMAIL_DOMAINS_URL}"
+            )
+            with httpx.Client(timeout=10.0) as client:
+                response = client.get(DISPOSABLE_EMAIL_DOMAINS_URL)
+                response.raise_for_status()
+
+                domains_list = response.json()
+
+                if not isinstance(domains_list, list):
+                    logger.error(
+                        f"Expected list from disposable domains URL, got {type(domains_list)}"
+                    )
+                    return self._fallback_domains.copy()
+
+                # Convert all to lowercase and create set
+                domains = {domain.lower().strip() for domain in domains_list if domain}
+
+                # Always include fallback domains
+                domains.update(self._fallback_domains)
+
+                logger.info(
+                    f"Successfully fetched {len(domains)} disposable email domains"
+                )
+                return domains
+
+        except httpx.HTTPError as e:
+            logger.warning(f"Failed to fetch disposable domains (HTTP error): {e}")
+        except Exception as e:
+            logger.warning(f"Failed to fetch disposable domains: {e}")
+
+        # On error, return fallback domains
+        return self._fallback_domains.copy()
+
+    def get_domains(self) -> Set[str]:
+        """
+        Get the cached set of disposable email domains.
+        Refreshes the cache if needed.
+
+        Returns:
+            Set of disposable domain strings (lowercased)
+        """
+        # Fast path: return cached domains if still fresh
+        if self._domains and not self._should_refresh():
+            return self._domains.copy()
+
+        # Slow path: need to refresh
+        with self._fetch_lock:
+            # Double-check after acquiring lock
+            if self._domains and not self._should_refresh():
+                return self._domains.copy()
+
+            self._domains = self._fetch_domains()
+            self._last_fetch_time = time.time()
+            return self._domains.copy()
+
+    def is_disposable(self, email: str) -> bool:
+        """
+        Check if an email address uses a disposable domain.
+
+        Args:
+            email: The email address to check
+
+        Returns:
+            True if the email domain is disposable, False otherwise
+        """
+        if not email or "@" not in email:
+            return False
+
+        parts = email.split("@")
+        if len(parts) != 2 or not parts[0]:  # Must have user@domain with non-empty user
+            return False
+
+        domain = parts[1].lower().strip()
+        if not domain:  # Domain part must not be empty
+            return False
+
+        disposable_domains = self.get_domains()
+        return domain in disposable_domains
+
+
+# Global singleton instance
+_validator = DisposableEmailValidator()
+
+
+def is_disposable_email(email: str) -> bool:
+    """
+    Check if an email address uses a disposable/temporary domain.
+
+    This is a convenience function that uses the global validator instance.
+
+    Args:
+        email: The email address to check
+
+    Returns:
+        True if the email uses a disposable domain, False otherwise
+    """
+    return _validator.is_disposable(email)
+
+
+def refresh_disposable_domains() -> None:
+    """
+    Force a refresh of the disposable domains list.
+
+    This can be called manually if you want to update the list
+    without waiting for the cache to expire.
+    """
+    _validator._last_fetch_time = 0
+    _validator.get_domains()

backend/onyx/auth/schemas.py

@@ -40,6 +40,8 @@ class UserRead(schemas.BaseUser[uuid.UUID]):
 class UserCreate(schemas.BaseUserCreate):
     role: UserRole = UserRole.BASIC
     tenant_id: str | None = None
+    # Captcha token for cloud signup protection (optional, only used when captcha is enabled)
+    captcha_token: str | None = None
 
 
 class UserUpdateWithRole(schemas.BaseUserUpdate):

backend/onyx/auth/users.py

@@ -60,6 +60,7 @@ from sqlalchemy.exc import IntegrityError
 from sqlalchemy.ext.asyncio import AsyncSession
 
 from onyx.auth.api_key import get_hashed_api_key_from_request
+from onyx.auth.disposable_email_validator import is_disposable_email
 from onyx.auth.email_utils import send_forgot_password_email
 from onyx.auth.email_utils import send_user_verification_email
 from onyx.auth.invited_users import get_invited_users
@@ -248,13 +249,23 @@ def verify_email_in_whitelist(email: str, tenant_id: str) -> None:
 
 
 def verify_email_domain(email: str) -> None:
+    if email.count("@") != 1:
+        raise HTTPException(
+            status_code=status.HTTP_400_BAD_REQUEST,
+            detail="Email is not valid",
+        )
+
+    domain = email.split("@")[-1].lower()
+
+    # Check if email uses a disposable/temporary domain
+    if is_disposable_email(email):
+        raise HTTPException(
+            status_code=status.HTTP_400_BAD_REQUEST,
+            detail="Disposable email addresses are not allowed. Please use a permanent email address.",
+        )
+
+    # Check domain whitelist if configured
     if VALID_EMAIL_DOMAINS:
-        if email.count("@") != 1:
-            raise HTTPException(
-                status_code=status.HTTP_400_BAD_REQUEST,
-                detail="Email is not valid",
-            )
-        domain = email.split("@")[-1].lower()
         if domain not in VALID_EMAIL_DOMAINS:
             raise HTTPException(
                 status_code=status.HTTP_400_BAD_REQUEST,
@@ -292,11 +303,57 @@ class UserManager(UUIDIDMixin, BaseUserManager[User, uuid.UUID]):
         safe: bool = False,
         request: Optional[Request] = None,
     ) -> User:
+        # Verify captcha if enabled (for cloud signup protection)
+        from onyx.auth.captcha import CaptchaVerificationError
+        from onyx.auth.captcha import is_captcha_enabled
+        from onyx.auth.captcha import verify_captcha_token
+
+        if is_captcha_enabled() and request is not None:
+            # Get captcha token from request body or headers
+            captcha_token = None
+            if hasattr(user_create, "captcha_token"):
+                captcha_token = getattr(user_create, "captcha_token", None)
+
+            # Also check headers as a fallback
+            if not captcha_token:
+                captcha_token = request.headers.get("X-Captcha-Token")
+
+            try:
+                await verify_captcha_token(
+                    captcha_token or "", expected_action="signup"
+                )
+            except CaptchaVerificationError as e:
+                raise HTTPException(
+                    status_code=status.HTTP_400_BAD_REQUEST,
+                    detail={"reason": str(e)},
+                )
+
         # We verify the password here to make sure it's valid before we proceed
         await self.validate_password(
             user_create.password, cast(schemas.UC, user_create)
         )
 
+        # Check for disposable emails BEFORE provisioning tenant
+        # This prevents creating tenants for throwaway email addresses
+        try:
+            verify_email_domain(user_create.email)
+        except HTTPException as e:
+            # Log blocked disposable email attempts
+            if (
+                e.status_code == status.HTTP_400_BAD_REQUEST
+                and "Disposable email" in str(e.detail)
+            ):
+                domain = (
+                    user_create.email.split("@")[-1]
+                    if "@" in user_create.email
+                    else "unknown"
+                )
+                logger.warning(
+                    f"Blocked disposable email registration attempt: {domain}",
+                    extra={"email_domain": domain},
+                )
+            raise
+
         user_count: int | None = None
         referral_source = (
             request.cookies.get("referral_source", None)
@@ -318,8 +375,17 @@ class UserManager(UUIDIDMixin, BaseUserManager[User, uuid.UUID]):
         token = CURRENT_TENANT_ID_CONTEXTVAR.set(tenant_id)
         try:
             async with get_async_session_context_manager(tenant_id) as db_session:
-                verify_email_is_invited(user_create.email)
-                verify_email_domain(user_create.email)
+                # Check invite list based on deployment mode
+                if MULTI_TENANT:
+                    # Multi-tenant: Only require invite for existing tenants
+                    # New tenant creation (first user) doesn't require an invite
+                    user_count = await get_user_count()
+                    if user_count > 0:
+                        # Tenant already has users - require invite for new users
+                        verify_email_is_invited(user_create.email)
+                else:
+                    # Single-tenant: Check invite list (skips if SAML/OIDC or no list configured)
+                    verify_email_is_invited(user_create.email)
                 if MULTI_TENANT:
                     tenant_user_db = SQLAlchemyUserAdminDB[User, uuid.UUID](
                         db_session, User, OAuthAccount

backend/onyx/background/celery/apps/app_base.py

@@ -26,6 +26,7 @@ from onyx.background.celery.celery_utils import celery_is_worker_primary
 from onyx.background.celery.celery_utils import make_probe_path
 from onyx.background.celery.tasks.vespa.document_sync import DOCUMENT_SYNC_PREFIX
 from onyx.background.celery.tasks.vespa.document_sync import DOCUMENT_SYNC_TASKSET_KEY
+from onyx.configs.app_configs import ENABLE_OPENSEARCH_FOR_ONYX
 from onyx.configs.constants import ONYX_CLOUD_CELERY_TASK_PREFIX
 from onyx.configs.constants import OnyxRedisLocks
 from onyx.db.engine.sql_engine import get_sqlalchemy_engine
@@ -515,6 +516,9 @@ def wait_for_vespa_or_shutdown(sender: Any, **kwargs: Any) -> None:
     """Waits for Vespa to become ready subject to a timeout.
     Raises WorkerShutdown if the timeout is reached."""
 
+    if ENABLE_OPENSEARCH_FOR_ONYX:
+        return
+
     if not wait_for_vespa_with_timeout():
         msg = "Vespa: Readiness probe did not succeed within the timeout. Exiting..."
         logger.error(msg)

backend/onyx/background/celery/apps/docfetching.py

@@ -98,8 +98,5 @@ for bootstep in base_bootsteps:
 celery_app.autodiscover_tasks(
     [
         "onyx.background.celery.tasks.docfetching",
-        # Ensure the user files indexing worker registers the doc_id migration task
-        # TODO(subash): remove this once the doc_id migration is complete
-        "onyx.background.celery.tasks.user_file_processing",
     ]
 )

backend/onyx/background/celery/tasks/beat_schedule.py

@@ -2,8 +2,12 @@ import copy
 from datetime import timedelta
 from typing import Any
 
+from celery.schedules import crontab
+
+from onyx.configs.app_configs import AUTO_LLM_CONFIG_URL
+from onyx.configs.app_configs import AUTO_LLM_UPDATE_INTERVAL_SECONDS
 from onyx.configs.app_configs import ENTERPRISE_EDITION_ENABLED
-from onyx.configs.app_configs import LLM_MODEL_UPDATE_API_URL
+from onyx.configs.app_configs import SCHEDULED_EVAL_DATASET_NAMES
 from onyx.configs.constants import ONYX_CLOUD_CELERY_TASK_PREFIX
 from onyx.configs.constants import OnyxCeleryPriority
 from onyx.configs.constants import OnyxCeleryQueues
@@ -53,16 +57,6 @@ beat_task_templates: list[dict] = [
             "expires": BEAT_EXPIRES_DEFAULT,
         },
     },
-    {
-        "name": "user-file-docid-migration",
-        "task": OnyxCeleryTask.USER_FILE_DOCID_MIGRATION,
-        "schedule": timedelta(minutes=10),
-        "options": {
-            "priority": OnyxCeleryPriority.HIGH,
-            "expires": BEAT_EXPIRES_DEFAULT,
-            "queue": OnyxCeleryQueues.USER_FILES_INDEXING,
-        },
-    },
     {
         "name": "check-for-kg-processing",
         "task": OnyxCeleryTask.CHECK_KG_PROCESSING,
@@ -171,13 +165,32 @@ if ENTERPRISE_EDITION_ENABLED:
         ]
     )
 
-# Only add the LLM model update task if the API URL is configured
-if LLM_MODEL_UPDATE_API_URL:
+# Add the Auto LLM update task if the config URL is set (has a default)
+if AUTO_LLM_CONFIG_URL:
     beat_task_templates.append(
         {
-            "name": "check-for-llm-model-update",
-            "task": OnyxCeleryTask.CHECK_FOR_LLM_MODEL_UPDATE,
-            "schedule": timedelta(hours=1),  # Check every hour
+            "name": "check-for-auto-llm-update",
+            "task": OnyxCeleryTask.CHECK_FOR_AUTO_LLM_UPDATE,
+            "schedule": timedelta(seconds=AUTO_LLM_UPDATE_INTERVAL_SECONDS),
+            "options": {
+                "priority": OnyxCeleryPriority.LOW,
+                "expires": AUTO_LLM_UPDATE_INTERVAL_SECONDS,
+            },
+        }
+    )
+
+# Add scheduled eval task if datasets are configured
+if SCHEDULED_EVAL_DATASET_NAMES:
+    beat_task_templates.append(
+        {
+            "name": "scheduled-eval-pipeline",
+            "task": OnyxCeleryTask.SCHEDULED_EVAL_TASK,
+            # run every Sunday at midnight UTC
+            "schedule": crontab(
+                hour=0,
+                minute=0,
+                day_of_week=0,
+            ),
             "options": {
                 "priority": OnyxCeleryPriority.LOW,
                 "expires": BEAT_EXPIRES_DEFAULT,

backend/onyx/background/celery/tasks/docfetching/task_creation_utils.py

@@ -72,15 +72,6 @@ def try_creating_docfetching_task(
             # Another indexing attempt is already running
             return None
 
-        # Determine which queue to use based on whether this is a user file
-        # TODO: at the moment the indexing pipeline is
-        # shared between user files and connectors
-        queue = (
-            OnyxCeleryQueues.USER_FILES_INDEXING
-            if cc_pair.is_user_file
-            else OnyxCeleryQueues.CONNECTOR_DOC_FETCHING
-        )
-
         # Use higher priority for first-time indexing to ensure new connectors
         # get processed before re-indexing of existing connectors
         has_successful_attempt = cc_pair.last_successful_index_time is not None
@@ -99,7 +90,7 @@ def try_creating_docfetching_task(
                 search_settings_id=search_settings.id,
                 tenant_id=tenant_id,
             ),
-            queue=queue,
+            queue=OnyxCeleryQueues.CONNECTOR_DOC_FETCHING,
             task_id=custom_task_id,
             priority=priority,
         )

backend/onyx/background/celery/tasks/docprocessing/tasks.py

@@ -12,6 +12,7 @@ from celery import Celery
 from celery import shared_task
 from celery import Task
 from celery.exceptions import SoftTimeLimitExceeded
+from fastapi import HTTPException
 from pydantic import BaseModel
 from redis import Redis
 from redis.lock import Lock as RedisLock
@@ -40,9 +41,11 @@ from onyx.background.indexing.checkpointing_utils import (
 )
 from onyx.background.indexing.index_attempt_utils import cleanup_index_attempts
 from onyx.background.indexing.index_attempt_utils import get_old_index_attempts
+from onyx.configs.app_configs import AUTH_TYPE
 from onyx.configs.app_configs import MANAGED_VESPA
 from onyx.configs.app_configs import VESPA_CLOUD_CERT_PATH
 from onyx.configs.app_configs import VESPA_CLOUD_KEY_PATH
+from onyx.configs.constants import AuthType
 from onyx.configs.constants import CELERY_GENERIC_BEAT_LOCK_TIMEOUT
 from onyx.configs.constants import CELERY_INDEXING_LOCK_TIMEOUT
 from onyx.configs.constants import MilestoneRecordType
@@ -59,11 +62,9 @@ from onyx.db.connector import mark_ccpair_with_indexing_trigger
 from onyx.db.connector_credential_pair import (
     fetch_indexable_standard_connector_credential_pair_ids,
 )
-from onyx.db.connector_credential_pair import (
-    fetch_indexable_user_file_connector_credential_pair_ids,
-)
 from onyx.db.connector_credential_pair import get_connector_credential_pair_from_id
 from onyx.db.connector_credential_pair import set_cc_pair_repeated_error_state
+from onyx.db.connector_credential_pair import update_connector_credential_pair_from_id
 from onyx.db.engine.sql_engine import get_session_with_current_tenant
 from onyx.db.engine.time_utils import get_db_current_time
 from onyx.db.enums import ConnectorCredentialPairStatus
@@ -96,9 +97,6 @@ from onyx.indexing.adapters.document_indexing_adapter import (
 from onyx.indexing.embedder import DefaultIndexingEmbedder
 from onyx.indexing.indexing_pipeline import run_indexing_pipeline
 from onyx.natural_language_processing.search_nlp_models import EmbeddingModel
-from onyx.natural_language_processing.search_nlp_models import (
-    InformationContentClassificationModel,
-)
 from onyx.natural_language_processing.search_nlp_models import warm_up_bi_encoder
 from onyx.redis.redis_connector import RedisConnector
 from onyx.redis.redis_pool import get_redis_client
@@ -115,6 +113,7 @@ from onyx.utils.telemetry import RecordType
 from shared_configs.configs import INDEXING_MODEL_SERVER_HOST
 from shared_configs.configs import INDEXING_MODEL_SERVER_PORT
 from shared_configs.configs import MULTI_TENANT
+from shared_configs.configs import USAGE_LIMITS_ENABLED
 from shared_configs.contextvars import CURRENT_TENANT_ID_CONTEXTVAR
 from shared_configs.contextvars import INDEX_ATTEMPT_INFO_CONTEXTVAR
 
@@ -541,12 +540,7 @@ def check_indexing_completion(
             ]:
                 # User file connectors must be paused on success
                 # NOTE: _run_indexing doesn't update connectors if the index attempt is the future embedding model
-                # TODO: figure out why this doesn't pause connectors during swap
-                cc_pair.status = (
-                    ConnectorCredentialPairStatus.PAUSED
-                    if cc_pair.is_user_file
-                    else ConnectorCredentialPairStatus.ACTIVE
-                )
+                cc_pair.status = ConnectorCredentialPairStatus.ACTIVE
                 db_session.commit()
 
             mt_cloud_telemetry(
@@ -812,13 +806,8 @@ def check_for_indexing(self: Task, *, tenant_id: str) -> int | None:
                     db_session, active_cc_pairs_only=True
                 )
             )
-            user_file_cc_pair_ids = (
-                fetch_indexable_user_file_connector_credential_pair_ids(
-                    db_session, search_settings_id=current_search_settings.id
-                )
-            )
 
-            primary_cc_pair_ids = standard_cc_pair_ids + user_file_cc_pair_ids
+            primary_cc_pair_ids = standard_cc_pair_ids
 
             # Get CC pairs for secondary search settings
             secondary_cc_pair_ids: list[int] = []
@@ -834,30 +823,47 @@ def check_for_indexing(self: Task, *, tenant_id: str) -> int | None:
                         db_session, active_cc_pairs_only=not include_paused
                     )
                 )
-                user_file_cc_pair_ids = (
-                    fetch_indexable_user_file_connector_credential_pair_ids(
-                        db_session, search_settings_id=secondary_search_settings.id
-                    )
-                    or []
-                )
 
-                secondary_cc_pair_ids = standard_cc_pair_ids + user_file_cc_pair_ids
+                secondary_cc_pair_ids = standard_cc_pair_ids
 
         # Flag CC pairs in repeated error state for primary/current search settings
         with get_session_with_current_tenant() as db_session:
             for cc_pair_id in primary_cc_pair_ids:
                 lock_beat.reacquire()
 
-                if is_in_repeated_error_state(
-                    cc_pair_id=cc_pair_id,
-                    search_settings_id=current_search_settings.id,
+                cc_pair = get_connector_credential_pair_from_id(
                     db_session=db_session,
+                    cc_pair_id=cc_pair_id,
+                )
+
+                # if already in repeated error state, don't do anything
+                # this is important so that we don't keep pausing the connector
+                # immediately upon a user un-pausing it to manually re-trigger and
+                # recover.
+                if (
+                    cc_pair
+                    and not cc_pair.in_repeated_error_state
+                    and is_in_repeated_error_state(
+                        cc_pair=cc_pair,
+                        search_settings_id=current_search_settings.id,
+                        db_session=db_session,
+                    )
                 ):
                     set_cc_pair_repeated_error_state(
                         db_session=db_session,
                         cc_pair_id=cc_pair_id,
                         in_repeated_error_state=True,
                     )
+                    # When entering repeated error state, also pause the connector
+                    # to prevent continued indexing retry attempts burning through embedding credits.
+                    # NOTE: only for Cloud, since most self-hosted users use self-hosted embedding
+                    # models. Also, they are more prone to repeated failures -> eventual success.
+                    if AUTH_TYPE == AuthType.CLOUD:
+                        update_connector_credential_pair_from_id(
+                            db_session=db_session,
+                            cc_pair_id=cc_pair.id,
+                            status=ConnectorCredentialPairStatus.PAUSED,
+                        )
 
         # NOTE: At this point, we haven't done heavy checks on whether or not the CC pairs should actually be indexed
         # Heavy check, should_index(), is called in _kickoff_indexing_tasks
@@ -1282,6 +1288,26 @@ def docprocessing_task(
         INDEX_ATTEMPT_INFO_CONTEXTVAR.reset(token)
 
 
+def _check_chunk_usage_limit(tenant_id: str) -> None:
+    """Check if chunk indexing usage limit has been exceeded.
+
+    Raises UsageLimitExceededError if the limit is exceeded.
+    """
+    if not USAGE_LIMITS_ENABLED:
+        return
+
+    from onyx.db.usage import UsageType
+    from onyx.server.usage_limits import check_usage_and_raise
+
+    with get_session_with_current_tenant() as db_session:
+        check_usage_and_raise(
+            db_session=db_session,
+            usage_type=UsageType.CHUNKS_INDEXED,
+            tenant_id=tenant_id,
+            pending_amount=0,  # Just check current usage
+        )
+
+
 def _docprocessing_task(
     index_attempt_id: int,
     cc_pair_id: int,
@@ -1293,6 +1319,25 @@ def _docprocessing_task(
     if tenant_id:
         CURRENT_TENANT_ID_CONTEXTVAR.set(tenant_id)
 
+    # Check if chunk indexing usage limit has been exceeded before processing
+    if USAGE_LIMITS_ENABLED:
+        try:
+            _check_chunk_usage_limit(tenant_id)
+        except HTTPException as e:
+            # Log the error and fail the indexing attempt
+            task_logger.error(
+                f"Chunk indexing usage limit exceeded for tenant {tenant_id}: {e}"
+            )
+            with get_session_with_current_tenant() as db_session:
+                from onyx.db.index_attempt import mark_attempt_failed
+
+                mark_attempt_failed(
+                    index_attempt_id=index_attempt_id,
+                    db_session=db_session,
+                    failure_reason=str(e),
+                )
+            raise
+
     task_logger.info(
         f"Processing document batch: "
         f"attempt={index_attempt_id} "
@@ -1391,10 +1436,6 @@ def _docprocessing_task(
                 callback=callback,
             )
 
-            information_content_classification_model = (
-                InformationContentClassificationModel()
-            )
-
             document_index = get_default_document_index(
                 index_attempt.search_settings,
                 None,
@@ -1432,7 +1473,6 @@ def _docprocessing_task(
             # real work happens here!
             index_pipeline_result = run_indexing_pipeline(
                 embedder=embedding_model,
-                information_content_classification_model=information_content_classification_model,
                 document_index=document_index,
                 ignore_time_skip=True,  # Documents are already filtered during extraction
                 db_session=db_session,
@@ -1442,6 +1482,23 @@ def _docprocessing_task(
                 adapter=adapter,
             )
 
+        # Track chunk indexing usage for cloud usage limits
+        if USAGE_LIMITS_ENABLED and index_pipeline_result.total_chunks > 0:
+            try:
+                from onyx.db.usage import increment_usage
+                from onyx.db.usage import UsageType
+
+                with get_session_with_current_tenant() as usage_db_session:
+                    increment_usage(
+                        db_session=usage_db_session,
+                        usage_type=UsageType.CHUNKS_INDEXED,
+                        amount=index_pipeline_result.total_chunks,
+                    )
+                    usage_db_session.commit()
+            except Exception as e:
+                # Log but don't fail indexing if usage tracking fails
+                task_logger.warning(f"Failed to track chunk indexing usage: {e}")
+
         # Update batch completion and document counts atomically using database coordination
 
         with get_session_with_current_tenant() as db_session, cross_batch_db_lock:

backend/onyx/background/celery/tasks/docprocessing/utils.py

@@ -10,7 +10,6 @@ from sqlalchemy.orm import Session
 from onyx.configs.app_configs import DISABLE_INDEX_UPDATE_ON_SWAP
 from onyx.configs.constants import CELERY_GENERIC_BEAT_LOCK_TIMEOUT
 from onyx.configs.constants import DocumentSource
-from onyx.db.connector_credential_pair import get_connector_credential_pair_from_id
 from onyx.db.engine.time_utils import get_db_current_time
 from onyx.db.enums import ConnectorCredentialPairStatus
 from onyx.db.enums import IndexingStatus
@@ -126,18 +125,9 @@ class IndexingCallback(IndexingHeartbeatInterface):
 
 
 def is_in_repeated_error_state(
-    cc_pair_id: int, search_settings_id: int, db_session: Session
+    cc_pair: ConnectorCredentialPair, search_settings_id: int, db_session: Session
 ) -> bool:
     """Checks if the cc pair / search setting combination is in a repeated error state."""
-    cc_pair = get_connector_credential_pair_from_id(
-        db_session=db_session,
-        cc_pair_id=cc_pair_id,
-    )
-    if not cc_pair:
-        raise RuntimeError(
-            f"is_in_repeated_error_state - could not find cc_pair with id={cc_pair_id}"
-        )
-
     # if the connector doesn't have a refresh_freq, a single failed attempt is enough
     number_of_failed_attempts_in_a_row_needed = (
         NUM_REPEAT_ERRORS_BEFORE_REPEATED_ERROR_STATE
@@ -146,7 +136,7 @@ def is_in_repeated_error_state(
     )
 
     most_recent_index_attempts = get_recent_attempts_for_cc_pair(
-        cc_pair_id=cc_pair_id,
+        cc_pair_id=cc_pair.id,
         search_settings_id=search_settings_id,
         limit=number_of_failed_attempts_in_a_row_needed,
         db_session=db_session,
@@ -180,7 +170,7 @@ def should_index(
         db_session=db_session,
     )
     all_recent_errored = is_in_repeated_error_state(
-        cc_pair_id=cc_pair.id,
+        cc_pair=cc_pair,
         search_settings_id=search_settings_instance.id,
         db_session=db_session,
     )

backend/onyx/background/celery/tasks/evals/tasks.py

@@ -1,9 +1,15 @@
+from datetime import datetime
+from datetime import timezone
 from typing import Any
 
 from celery import shared_task
 from celery import Task
 
+from onyx.configs.app_configs import BRAINTRUST_API_KEY
 from onyx.configs.app_configs import JOB_TIMEOUT
+from onyx.configs.app_configs import SCHEDULED_EVAL_DATASET_NAMES
+from onyx.configs.app_configs import SCHEDULED_EVAL_PERMISSIONS_EMAIL
+from onyx.configs.app_configs import SCHEDULED_EVAL_PROJECT
 from onyx.configs.constants import OnyxCeleryTask
 from onyx.evals.eval import run_eval
 from onyx.evals.models import EvalConfigurationOptions
@@ -33,3 +39,109 @@ def eval_run_task(
     except Exception:
         logger.error("Failed to run eval task")
         raise
+
+
+@shared_task(
+    name=OnyxCeleryTask.SCHEDULED_EVAL_TASK,
+    ignore_result=True,
+    soft_time_limit=JOB_TIMEOUT * 5,  # Allow more time for multiple datasets
+    bind=True,
+    trail=False,
+)
+def scheduled_eval_task(self: Task, **kwargs: Any) -> None:
+    """
+    Scheduled task to run evaluations on configured datasets.
+    Runs weekly on Sunday at midnight UTC.
+
+    Configure via environment variables (with defaults):
+    - SCHEDULED_EVAL_DATASET_NAMES: Comma-separated list of Braintrust dataset names
+    - SCHEDULED_EVAL_PERMISSIONS_EMAIL: Email for search permissions (default: roshan@onyx.app)
+    - SCHEDULED_EVAL_PROJECT: Braintrust project name
+    """
+    if not BRAINTRUST_API_KEY:
+        logger.error("BRAINTRUST_API_KEY is not configured, cannot run scheduled evals")
+        return
+
+    if not SCHEDULED_EVAL_PROJECT:
+        logger.error(
+            "SCHEDULED_EVAL_PROJECT is not configured, cannot run scheduled evals"
+        )
+        return
+
+    if not SCHEDULED_EVAL_DATASET_NAMES:
+        logger.info("No scheduled eval datasets configured, skipping")
+        return
+
+    if not SCHEDULED_EVAL_PERMISSIONS_EMAIL:
+        logger.error("SCHEDULED_EVAL_PERMISSIONS_EMAIL not configured")
+        return
+
+    project_name = SCHEDULED_EVAL_PROJECT
+    dataset_names = SCHEDULED_EVAL_DATASET_NAMES
+    permissions_email = SCHEDULED_EVAL_PERMISSIONS_EMAIL
+
+    # Create a timestamp for the scheduled run
+    run_timestamp = datetime.now(timezone.utc).strftime("%Y-%m-%d")
+
+    logger.info(
+        f"Starting scheduled eval pipeline for project '{project_name}' "
+        f"with {len(dataset_names)} dataset(s): {dataset_names}"
+    )
+
+    pipeline_start = datetime.now(timezone.utc)
+    results: list[dict[str, Any]] = []
+
+    for dataset_name in dataset_names:
+        start_time = datetime.now(timezone.utc)
+        error_message: str | None = None
+        success = False
+
+        # Create informative experiment name for scheduled runs
+        experiment_name = f"{dataset_name} - {run_timestamp}"
+
+        try:
+            logger.info(
+                f"Running scheduled eval for dataset: {dataset_name} "
+                f"(project: {project_name})"
+            )
+
+            configuration = EvalConfigurationOptions(
+                search_permissions_email=permissions_email,
+                dataset_name=dataset_name,
+                no_send_logs=False,
+                braintrust_project=project_name,
+                experiment_name=experiment_name,
+            )
+
+            result = run_eval(
+                configuration=configuration,
+                remote_dataset_name=dataset_name,
+            )
+            success = result.success
+            logger.info(f"Completed eval for {dataset_name}: success={success}")
+
+        except Exception as e:
+            logger.exception(f"Failed to run scheduled eval for {dataset_name}")
+            error_message = str(e)
+            success = False
+
+        end_time = datetime.now(timezone.utc)
+
+        results.append(
+            {
+                "dataset_name": dataset_name,
+                "success": success,
+                "start_time": start_time,
+                "end_time": end_time,
+                "error_message": error_message,
+            }
+        )
+
+    pipeline_end = datetime.now(timezone.utc)
+    total_duration = (pipeline_end - pipeline_start).total_seconds()
+
+    passed_count = sum(1 for r in results if r["success"])
+    logger.info(
+        f"Scheduled eval pipeline completed: {passed_count}/{len(results)} passed "
+        f"in {total_duration:.1f}s"
+    )

backend/onyx/background/celery/tasks/llm_model_update/tasks.py

@@ -1,135 +1,57 @@
-from typing import Any
-
-import requests
 from celery import shared_task
 from celery import Task
 
 from onyx.background.celery.apps.app_base import task_logger
-from onyx.configs.app_configs import JOB_TIMEOUT
-from onyx.configs.app_configs import LLM_MODEL_UPDATE_API_URL
+from onyx.configs.app_configs import AUTO_LLM_CONFIG_URL
 from onyx.configs.constants import OnyxCeleryTask
 from onyx.db.engine.sql_engine import get_session_with_current_tenant
-from onyx.db.models import LLMProvider
-from onyx.db.models import ModelConfiguration
-
-
-def _process_model_list_response(model_list_json: Any) -> list[str]:
-    # Handle case where response is wrapped in a "data" field
-    if isinstance(model_list_json, dict):
-        if "data" in model_list_json:
-            model_list_json = model_list_json["data"]
-        elif "models" in model_list_json:
-            model_list_json = model_list_json["models"]
-        else:
-            raise ValueError(
-                "Invalid response from API - expected dict with 'data' or "
-                f"'models' field, got {type(model_list_json)}"
-            )
-
-    if not isinstance(model_list_json, list):
-        raise ValueError(
-            f"Invalid response from API - expected list, got {type(model_list_json)}"
-        )
-
-    # Handle both string list and object list cases
-    model_names: list[str] = []
-    for item in model_list_json:
-        if isinstance(item, str):
-            model_names.append(item)
-        elif isinstance(item, dict):
-            if "model_name" in item:
-                model_names.append(item["model_name"])
-            elif "id" in item:
-                model_names.append(item["id"])
-            else:
-                raise ValueError(
-                    f"Invalid item in model list - expected dict with model_name or id, got {type(item)}"
-                )
-        else:
-            raise ValueError(
-                f"Invalid item in model list - expected string or dict, got {type(item)}"
-            )
-
-    return model_names
 
 
 @shared_task(
-    name=OnyxCeleryTask.CHECK_FOR_LLM_MODEL_UPDATE,
+    name=OnyxCeleryTask.CHECK_FOR_AUTO_LLM_UPDATE,
     ignore_result=True,
-    soft_time_limit=JOB_TIMEOUT,
+    soft_time_limit=300,  # 5 minute timeout
     trail=False,
     bind=True,
 )
-def check_for_llm_model_update(self: Task, *, tenant_id: str) -> bool | None:
-    if not LLM_MODEL_UPDATE_API_URL:
-        raise ValueError("LLM model update API URL not configured")
+def check_for_auto_llm_updates(self: Task, *, tenant_id: str) -> bool | None:
+    """Periodic task to fetch LLM model updates from GitHub
+    and sync them to providers in Auto mode.
 
-    # First fetch the models from the API
-    try:
-        response = requests.get(LLM_MODEL_UPDATE_API_URL)
-        response.raise_for_status()
-        available_models = _process_model_list_response(response.json())
-        task_logger.info(f"Found available models: {available_models}")
-
-    except Exception:
-        task_logger.exception("Failed to fetch models from API.")
+    This task checks the GitHub-hosted config file and updates all
+    providers that have is_auto_mode=True.
+    """
+    if not AUTO_LLM_CONFIG_URL:
+        task_logger.debug("AUTO_LLM_CONFIG_URL not configured, skipping")
         return None
 
-    # Then update the database with the fetched models
-    with get_session_with_current_tenant() as db_session:
-        # Get the default LLM provider
-        default_provider = (
-            db_session.query(LLMProvider)
-            .filter(LLMProvider.is_default_provider.is_(True))
-            .first()
+    try:
+        # Import here to avoid circular imports
+        from onyx.llm.well_known_providers.auto_update_service import (
+            fetch_llm_recommendations_from_github,
+        )
+        from onyx.llm.well_known_providers.auto_update_service import (
+            sync_llm_models_from_github,
         )
 
-        if not default_provider:
-            task_logger.warning("No default LLM provider found")
+        # Fetch config from GitHub
+        config = fetch_llm_recommendations_from_github()
+
+        if not config:
+            task_logger.warning("Failed to fetch GitHub config")
             return None
 
-        # log change if any
-        old_models = set(
-            model_configuration.name
-            for model_configuration in default_provider.model_configurations
-        )
-        new_models = set(available_models)
-        added_models = new_models - old_models
-        removed_models = old_models - new_models
+        # Sync to database
+        with get_session_with_current_tenant() as db_session:
+            results = sync_llm_models_from_github(db_session, config)
 
-        if added_models:
-            task_logger.info(f"Adding models: {sorted(added_models)}")
-        if removed_models:
-            task_logger.info(f"Removing models: {sorted(removed_models)}")
+            if results:
+                task_logger.info(f"Auto mode sync results: {results}")
+            else:
+                task_logger.debug("No model updates applied")
 
-        # Update the provider's model list
-        # Remove models that are no longer available
-        db_session.query(ModelConfiguration).filter(
-            ModelConfiguration.llm_provider_id == default_provider.id,
-            ModelConfiguration.name.notin_(available_models),
-        ).delete(synchronize_session=False)
-
-        # Add new models
-        for available_model_name in available_models:
-            db_session.merge(
-                ModelConfiguration(
-                    llm_provider_id=default_provider.id,
-                    name=available_model_name,
-                    is_visible=False,
-                    max_input_tokens=None,
-                )
-            )
-
-        # if the default model is no longer available, set it to the first model in the list
-        if default_provider.default_model_name not in available_models:
-            task_logger.info(
-                f"Default model {default_provider.default_model_name} not "
-                f"available, setting to first model in list: {available_models[0]}"
-            )
-            default_provider.default_model_name = available_models[0]
-        db_session.commit()
-
-        if added_models or removed_models:
-            task_logger.info("Updated model list for default provider.")
+    except Exception:
+        task_logger.exception("Error in auto LLM update task")
+        raise
 
     return True

backend/onyx/background/celery/tasks/monitoring/tasks.py

@@ -886,9 +886,7 @@ def monitor_celery_queues_helper(
         OnyxCeleryQueues.CONNECTOR_DOC_FETCHING, r_celery
     )
     n_docprocessing = celery_get_queue_length(OnyxCeleryQueues.DOCPROCESSING, r_celery)
-    n_user_files_indexing = celery_get_queue_length(
-        OnyxCeleryQueues.USER_FILES_INDEXING, r_celery
-    )
+
     n_user_file_processing = celery_get_queue_length(
         OnyxCeleryQueues.USER_FILE_PROCESSING, r_celery
     )
@@ -924,7 +922,6 @@ def monitor_celery_queues_helper(
         f"docfetching_prefetched={len(n_docfetching_prefetched)} "
         f"docprocessing={n_docprocessing} "
         f"docprocessing_prefetched={len(n_docprocessing_prefetched)} "
-        f"user_files_indexing={n_user_files_indexing} "
         f"user_file_processing={n_user_file_processing} "
         f"user_file_project_sync={n_user_file_project_sync} "
         f"user_file_delete={n_user_file_delete} "

backend/onyx/background/celery/tasks/user_file_processing/tasks.py

@@ -1,6 +1,5 @@
 import datetime
 import time
-from collections.abc import Sequence
 from typing import Any
 from uuid import UUID
 
@@ -19,11 +18,9 @@ from onyx.configs.app_configs import MANAGED_VESPA
 from onyx.configs.app_configs import VESPA_CLOUD_CERT_PATH
 from onyx.configs.app_configs import VESPA_CLOUD_KEY_PATH
 from onyx.configs.constants import CELERY_GENERIC_BEAT_LOCK_TIMEOUT
-from onyx.configs.constants import CELERY_USER_FILE_DOCID_MIGRATION_LOCK_TIMEOUT
 from onyx.configs.constants import CELERY_USER_FILE_PROCESSING_LOCK_TIMEOUT
 from onyx.configs.constants import CELERY_USER_FILE_PROJECT_SYNC_LOCK_TIMEOUT
 from onyx.configs.constants import DocumentSource
-from onyx.configs.constants import FileOrigin
 from onyx.configs.constants import OnyxCeleryPriority
 from onyx.configs.constants import OnyxCeleryQueues
 from onyx.configs.constants import OnyxCeleryTask
@@ -32,28 +29,18 @@ from onyx.connectors.file.connector import LocalFileConnector
 from onyx.connectors.models import Document
 from onyx.db.engine.sql_engine import get_session_with_current_tenant
 from onyx.db.enums import UserFileStatus
-from onyx.db.models import FileRecord
-from onyx.db.models import SearchDoc
 from onyx.db.models import UserFile
 from onyx.db.search_settings import get_active_search_settings
 from onyx.db.search_settings import get_active_search_settings_list
 from onyx.document_index.factory import get_default_document_index
-from onyx.document_index.interfaces import VespaDocumentFields
 from onyx.document_index.interfaces import VespaDocumentUserFields
-from onyx.document_index.vespa.shared_utils.utils import (
-    replace_invalid_doc_id_characters,
-)
 from onyx.document_index.vespa_constants import DOCUMENT_ID_ENDPOINT
 from onyx.file_store.file_store import get_default_file_store
-from onyx.file_store.file_store import S3BackedFileStore
 from onyx.file_store.utils import user_file_id_to_plaintext_file_name
 from onyx.httpx.httpx_pool import HttpxPool
 from onyx.indexing.adapters.user_file_indexing_adapter import UserFileIndexingAdapter
 from onyx.indexing.embedder import DefaultIndexingEmbedder
 from onyx.indexing.indexing_pipeline import run_indexing_pipeline
-from onyx.natural_language_processing.search_nlp_models import (
-    InformationContentClassificationModel,
-)
 from onyx.redis.redis_pool import get_redis_client
 
 
@@ -257,10 +244,6 @@ def process_single_user_file(self: Task, *, user_file_id: str, tenant_id: str) -
                     search_settings=current_search_settings,
                 )
 
-                information_content_classification_model = (
-                    InformationContentClassificationModel()
-                )
-
                 document_index = get_default_document_index(
                     current_search_settings,
                     None,
@@ -275,7 +258,6 @@ def process_single_user_file(self: Task, *, user_file_id: str, tenant_id: str) -
                 # real work happens here!
                 index_pipeline_result = run_indexing_pipeline(
                     embedder=embedding_model,
-                    information_content_classification_model=information_content_classification_model,
                     document_index=document_index,
                     ignore_time_skip=True,
                     db_session=db_session,
@@ -626,315 +608,3 @@ def process_single_user_file_project_sync(
             file_lock.release()
 
     return None
-
-
-def _normalize_legacy_user_file_doc_id(old_id: str) -> str:
-    # Convert USER_FILE_CONNECTOR__<uuid> -> FILE_CONNECTOR__<uuid> for legacy values
-    user_prefix = "USER_FILE_CONNECTOR__"
-    file_prefix = "FILE_CONNECTOR__"
-    if old_id.startswith(user_prefix):
-        remainder = old_id[len(user_prefix) :]
-        return file_prefix + remainder
-    return old_id
-
-
-def update_legacy_plaintext_file_records() -> None:
-    """Migrate legacy plaintext cache objects from int-based keys to UUID-based
-    keys. Copies each S3 object to its expected UUID key and updates DB.
-
-    Examples:
-    - Old key: bucket/schema/plaintext_<int>
-    - New key: bucket/schema/plaintext_<uuid>
-    """
-
-    task_logger.info("update_legacy_plaintext_file_records - Starting")
-
-    with get_session_with_current_tenant() as db_session:
-        store = get_default_file_store()
-
-        if not isinstance(store, S3BackedFileStore):
-            task_logger.info(
-                "update_legacy_plaintext_file_records - Skipping non-S3 store"
-            )
-            return
-
-        s3_client = store._get_s3_client()
-        bucket_name = store._get_bucket_name()
-
-        # Select PLAINTEXT_CACHE records whose object_key ends with 'plaintext_' + non-hyphen chars
-        # Example: 'some/path/plaintext_abc123' matches; '.../plaintext_foo-bar' does not
-        plaintext_records: Sequence[FileRecord] = (
-            db_session.execute(
-                sa.select(FileRecord).where(
-                    FileRecord.file_origin == FileOrigin.PLAINTEXT_CACHE,
-                    FileRecord.object_key.op("~")(r"plaintext_[^-]+$"),
-                )
-            )
-            .scalars()
-            .all()
-        )
-
-        task_logger.info(
-            f"update_legacy_plaintext_file_records - Found {len(plaintext_records)} plaintext records to update"
-        )
-
-        normalized = 0
-        for fr in plaintext_records:
-            try:
-                expected_key = store._get_s3_key(fr.file_id)
-                if fr.object_key == expected_key:
-                    continue
-
-                if fr.bucket_name is None:
-                    task_logger.warning(f"id={fr.file_id} - Bucket name is None")
-                    continue
-
-                if fr.object_key is None:
-                    task_logger.warning(f"id={fr.file_id} - Object key is None")
-                    continue
-
-                # Copy old object to new key
-                copy_source = f"{fr.bucket_name}/{fr.object_key}"
-                s3_client.copy_object(
-                    CopySource=copy_source,
-                    Bucket=bucket_name,
-                    Key=expected_key,
-                    MetadataDirective="COPY",
-                )
-
-                # Delete old object (best-effort)
-                try:
-                    s3_client.delete_object(Bucket=fr.bucket_name, Key=fr.object_key)
-                except Exception:
-                    pass
-
-                # Update DB record with new key
-                fr.object_key = expected_key
-                db_session.add(fr)
-                normalized += 1
-            except Exception as e:
-                task_logger.warning(f"id={fr.file_id} - {e.__class__.__name__}")
-
-        if normalized:
-            db_session.commit()
-            task_logger.info(
-                f"user_file_docid_migration_task normalized {normalized} plaintext objects"
-            )
-
-
-@shared_task(
-    name=OnyxCeleryTask.USER_FILE_DOCID_MIGRATION,
-    ignore_result=True,
-    bind=True,
-)
-def user_file_docid_migration_task(self: Task, *, tenant_id: str) -> bool:
-
-    task_logger.info(
-        f"user_file_docid_migration_task - Starting for tenant={tenant_id}"
-    )
-
-    redis_client = get_redis_client(tenant_id=tenant_id)
-    lock: RedisLock = redis_client.lock(
-        OnyxRedisLocks.USER_FILE_DOCID_MIGRATION_LOCK,
-        timeout=CELERY_USER_FILE_DOCID_MIGRATION_LOCK_TIMEOUT,
-    )
-
-    if not lock.acquire(blocking=False):
-        task_logger.info(
-            f"user_file_docid_migration_task - Lock held, skipping tenant={tenant_id}"
-        )
-        return False
-
-    updated_count = 0
-    try:
-        update_legacy_plaintext_file_records()
-        # Track lock renewal
-        last_lock_time = time.monotonic()
-        with get_session_with_current_tenant() as db_session:
-
-            # 20 is the documented default for httpx max_keepalive_connections
-            if MANAGED_VESPA:
-                httpx_init_vespa_pool(
-                    20, ssl_cert=VESPA_CLOUD_CERT_PATH, ssl_key=VESPA_CLOUD_KEY_PATH
-                )
-            else:
-                httpx_init_vespa_pool(20)
-
-            active_settings = get_active_search_settings(db_session)
-            document_index = get_default_document_index(
-                search_settings=active_settings.primary,
-                secondary_search_settings=active_settings.secondary,
-                httpx_client=HttpxPool.get("vespa"),
-            )
-
-            retry_index = RetryDocumentIndex(document_index)
-
-            # Select user files with a legacy doc id that have not been migrated
-            user_files = (
-                db_session.execute(
-                    sa.select(UserFile).where(
-                        sa.and_(
-                            UserFile.document_id.is_not(None),
-                            UserFile.document_id_migrated.is_(False),
-                        )
-                    )
-                )
-                .scalars()
-                .all()
-            )
-
-            task_logger.info(
-                f"user_file_docid_migration_task - Found {len(user_files)} user files to migrate"
-            )
-
-            # Query all SearchDocs that need updating
-            search_docs = (
-                db_session.execute(
-                    sa.select(SearchDoc).where(
-                        SearchDoc.document_id.like("%FILE_CONNECTOR__%")
-                    )
-                )
-                .scalars()
-                .all()
-            )
-            task_logger.info(
-                f"user_file_docid_migration_task - Found {len(search_docs)} search docs to update"
-            )
-
-            # Build a map of normalized doc IDs to SearchDocs
-            search_doc_map: dict[str, list[SearchDoc]] = {}
-            for sd in search_docs:
-                doc_id = sd.document_id
-                if search_doc_map.get(doc_id) is None:
-                    search_doc_map[doc_id] = []
-                search_doc_map[doc_id].append(sd)
-
-            task_logger.debug(
-                f"user_file_docid_migration_task - Built search doc map with {len(search_doc_map)} entries"
-            )
-
-            ids_preview = list(search_doc_map.keys())[:5]
-            task_logger.debug(
-                f"user_file_docid_migration_task - First few search_doc_map ids: {ids_preview if ids_preview else 'No ids found'}"
-            )
-            task_logger.debug(
-                f"user_file_docid_migration_task - search_doc_map total items: "
-                f"{sum(len(docs) for docs in search_doc_map.values())}"
-            )
-            for user_file in user_files:
-                # Periodically renew the Redis lock to prevent expiry mid-run
-                current_time = time.monotonic()
-                if current_time - last_lock_time >= (
-                    CELERY_USER_FILE_DOCID_MIGRATION_LOCK_TIMEOUT / 4
-                ):
-                    renewed = False
-                    try:
-                        # extend lock ttl to full timeout window
-                        lock.extend(CELERY_USER_FILE_DOCID_MIGRATION_LOCK_TIMEOUT)
-                        renewed = True
-                    except Exception:
-                        # if extend fails, best-effort reacquire as a fallback
-                        try:
-                            lock.reacquire()
-                            renewed = True
-                        except Exception:
-                            renewed = False
-                    last_lock_time = current_time
-                    if not renewed or not lock.owned():
-                        task_logger.error(
-                            "user_file_docid_migration_task - Lost lock ownership or failed to renew; aborting for safety"
-                        )
-                        return False
-
-                try:
-                    clean_old_doc_id = replace_invalid_doc_id_characters(
-                        user_file.document_id
-                    )
-                    normalized_doc_id = _normalize_legacy_user_file_doc_id(
-                        clean_old_doc_id
-                    )
-                    user_project_ids = [project.id for project in user_file.projects]
-                    task_logger.info(
-                        f"user_file_docid_migration_task - Migrating user file {user_file.id} with doc_id {normalized_doc_id}"
-                    )
-
-                    index_name = active_settings.primary.index_name
-
-                    # First find the chunks count using direct Vespa query
-                    selection = f"{index_name}.document_id=='{normalized_doc_id}'"
-
-                    # Count all chunks for this document
-                    chunk_count = _get_document_chunk_count(
-                        index_name=index_name,
-                        selection=selection,
-                    )
-
-                    task_logger.info(
-                        f"Found {chunk_count} chunks for document {normalized_doc_id}"
-                    )
-
-                    # Now update Vespa chunks with the found chunk count using retry_index
-                    # WARNING: In the future this will error; we no longer want
-                    # to support changing document ID.
-                    # TODO(andrei): Delete soon.
-                    retry_index.update_single(
-                        doc_id=str(normalized_doc_id),
-                        tenant_id=tenant_id,
-                        chunk_count=chunk_count,
-                        fields=VespaDocumentFields(document_id=str(user_file.id)),
-                        user_fields=VespaDocumentUserFields(
-                            user_projects=user_project_ids
-                        ),
-                    )
-                    user_file.chunk_count = chunk_count
-
-                    # Update the SearchDocs
-                    actual_doc_id = str(user_file.document_id)
-                    normalized_actual_doc_id = _normalize_legacy_user_file_doc_id(
-                        actual_doc_id
-                    )
-                    if (
-                        normalized_doc_id in search_doc_map
-                        or normalized_actual_doc_id in search_doc_map
-                    ):
-                        to_update = (
-                            search_doc_map[normalized_doc_id]
-                            if normalized_doc_id in search_doc_map
-                            else search_doc_map[normalized_actual_doc_id]
-                        )
-                        task_logger.debug(
-                            f"user_file_docid_migration_task - Updating {len(to_update)} search docs for user file {user_file.id}"
-                        )
-                        for search_doc in to_update:
-                            search_doc.document_id = str(user_file.id)
-                            db_session.add(search_doc)
-
-                    user_file.document_id_migrated = True
-                    db_session.add(user_file)
-                    db_session.commit()
-                    updated_count += 1
-                except Exception as per_file_exc:
-                    # Rollback the current transaction and continue with the next file
-                    db_session.rollback()
-                    task_logger.exception(
-                        f"user_file_docid_migration_task - Error migrating user file {user_file.id} - "
-                        f"{per_file_exc.__class__.__name__}"
-                    )
-
-            task_logger.info(
-                f"user_file_docid_migration_task - Updated {updated_count} user files"
-            )
-
-        task_logger.info(
-            f"user_file_docid_migration_task - Completed for tenant={tenant_id} (updated={updated_count})"
-        )
-        return True
-    except Exception as e:
-        task_logger.exception(
-            f"user_file_docid_migration_task - Error during execution for tenant={tenant_id} "
-            f"(updated={updated_count}) exception={e.__class__.__name__}"
-        )
-        return False
-    finally:
-        if lock.owned():
-            lock.release()

backend/onyx/chat/README.md

@@ -63,7 +63,7 @@ To ensure the LLM follows certain specific instructions, instructions are added
 tool is used, a citation reminder is always added. Otherwise, by default there is no reminder. If the user configures reminders, those are added to the
 final message. If a search related tool just ran and the user has reminders, both appear in a single message.
 
-If a search related tool is called at any point during the turn, the reminder will remain at the end until the turn is over and the agent as responded.
+If a search related tool is called at any point during the turn, the reminder will remain at the end until the turn is over and the agent has responded.
 
 
 ## Tool Calls
@@ -145,9 +145,83 @@ attention despite having global access.
 In a similar concept, LLM instructions in the system prompt are structured specifically so that there are coherent sections for the LLM to attend to. This is
 fairly surprising actually but if there is a line of instructions effectively saying "If you try to use some tools and find that you need more information or
 need to call additional tools, you are encouraged to do this", having this in the Tool section of the System prompt makes all the LLMs follow it well but if it's
-even just a paragraph away like near the beginning of the prompt, it is often often ignored. The difference is as drastic as a 30% follow rate to a 90% follow
+even just a paragraph away like near the beginning of the prompt, it is often ignored. The difference is as drastic as a 30% follow rate to a 90% follow
 rate even just moving the same statement a few sentences.
 
 
 ## Other related pointers
 - How messages, files, images are stored can be found in backend/onyx/db/models.py, there is also a README.md under that directory that may be helpful.
+
+---
+
+# Overview of LLM flow architecture
+
+**Concepts:**
+Turn: User sends a message and AI does some set of things and responds
+Step/Cycle: 1 single LLM inference given some context and some tools
+
+
+## 1. Top Level (process_message function):
+This function can be thought of as the set-up and validation layer. It ensures that the database is in a valid state, reads the
+messages in the session and sets up all the necessary items to run the chat loop and state containers. The major things it does
+are:
+- Validates the request
+- Builds the chat history for the session
+- Fetches any additional context such as files and images
+- Prepares all of the tools for the LLM
+- Creates the state container objects for use in the loop
+
+### Wrapper (run_chat_loop_with_state_containers function):
+This wrapper is used to run the LLM flow in a background thread and monitor the emitter for stop signals. This means the top
+level is as isolated from the LLM flow as possible and can continue to yield packets as soon as they are available from the lower
+levels. This also means that if the lower levels fail, the top level will still guarantee a reasonable response to the user.
+All of the saving and database operations are abstracted away from the lower levels.
+
+### Emitter
+The emitter is designed to be an object queue so that lower levels do not need to yield objects all the way back to the top.
+This way the functions can be better designed (not everything as a generator) and more easily tested. The wrapper around the
+LLM flow (run_chat_loop_with_state_containers) is used to monitor the emitter and handle packets as soon as they are available
+from the lower levels. Both the emitter and the state container are mutating state objects and only used to accumulate state.
+There should be no logic dependent on the states of these objects, especially in the lower levels. The emitter should only take
+packets and should not be used for other things.
+
+### State Container
+The state container is used to accumulate state during the LLM flow. Similar to the emitter, it should not be used for logic,
+only for accumulating state. It is used to gather all of the necessary information for saving the chat turn into the database.
+So it will accumulate answer tokens, reasoning tokens, tool calls, citation info, etc. This is used at the end of the flow once
+the lower level is completed whether on its own or stopped by the user. At that point, all of the state is read and stored into
+the database. The state container can be added to by any of the underlying layers, this is fine.
+
+### Stopping Generation
+A stop signal is checked every 300ms by the wrapper around the LLM flow. The signal itself
+is stored in Redis and is set by the user calling the stop endpoint. The wrapper ensures that no matter what the lower level is
+doing at the time, the thread can be killed by the top level. It does not require a cooperative cancellation from the lower level
+and in fact the lower level does not know about the stop signal at all.
+
+
+## 2. LLM Loop (run_llm_loop function)
+This function handles the logic of the Turn. It's essentially a while loop where context is added and modified (according what
+is outlined in the first half of this doc). Its main functionality is:
+- Translate and truncate the context for the LLM inference
+- Add context modifiers like reminders, updates to the system prompts, etc.
+- Run tool calls and gather results
+- Build some of the objects stored in the state container.
+
+
+## 3. LLM Step (run_llm_step function)
+This function is a single inference of the LLM. It's a wrapper around the LLM stream function which handles packet translations
+so that the Emitter can emit individual tokens as soon as they arrive. It also keeps track of the different sections since they
+do not all come at once (reasoning, answers, tool calls are all built up token by token). This layer also tracks the different
+tool calls and returns that to the LLM Loop to execute.
+
+
+## Things to know
+- Packets are labeled with a "turn_index" field as part of the Placement of the packet. This is not the same as the backend
+concept of a turn. The turn_index for the frontend is which block does this packet belong to. So while a reasoning + tool call
+comes from the same LLM inference (same backend LLM step), they are 2 turns to the frontend because that's how it's rendered.
+
+- There are 3 representations of "message". The first is the database model ChatMessage, this one should be translated away and
+not used deep into the flow. The second is ChatMessageSimple which is the data model which should be used throughout the code
+as much as possible. If modifications/additions are needed, it should be to this object. This is the rich representation of a
+message for the code. Finally there is the LanguageModelInput representation of a message. This one is for the LLM interface
+layer and is as stripped down as possible so that the LLM interface can be clean and easy to maintain/extend.

backend/onyx/chat/chat_state.py

@@ -1,4 +1,5 @@
 import threading
+import time
 from collections.abc import Callable
 from collections.abc import Generator
 from queue import Empty
@@ -38,6 +39,7 @@ class ChatStateContainer:
         self.citation_to_doc: CitationMapping = {}
         # True if this turn is a clarification question (deep research flow)
         self.is_clarification: bool = False
+        # Note: LLM cost tracking is now handled in multi_llm.py
 
     def add_tool_call(self, tool_call: ToolCallInfo) -> None:
         """Add a tool call to the accumulated state."""
@@ -144,6 +146,9 @@ def run_chat_loop_with_state_containers(
     thread = run_in_background(run_with_exception_capture)
 
     pkt: Packet | None = None
+    last_turn_index = 0  # Track the highest turn_index seen for stop packet
+    last_cancel_check = time.monotonic()
+    cancel_check_interval = 0.3  # Check for cancellation every 300ms
     try:
         while True:
             # Poll queue with 300ms timeout for natural stop signal checking
@@ -152,18 +157,40 @@ def run_chat_loop_with_state_containers(
                 pkt = emitter.bus.get(timeout=0.3)
             except Empty:
                 if not is_connected():
-                    # Stop signal detected, kill the thread
+                    # Stop signal detected
+                    yield Packet(
+                        placement=Placement(turn_index=last_turn_index + 1),
+                        obj=OverallStop(type="stop", stop_reason="user_cancelled"),
+                    )
                     break
+                last_cancel_check = time.monotonic()
                 continue
 
             if pkt is not None:
-                if pkt.obj == OverallStop(type="stop"):
+                # Track the highest turn_index for the stop packet
+                if pkt.placement and pkt.placement.turn_index > last_turn_index:
+                    last_turn_index = pkt.placement.turn_index
+
+                if isinstance(pkt.obj, OverallStop):
                     yield pkt
                     break
                 elif isinstance(pkt.obj, PacketException):
                     raise pkt.obj.exception
                 else:
                     yield pkt
+
+                # Check for cancellation periodically even when packets are flowing
+                # This ensures stop signal is checked during active streaming
+                current_time = time.monotonic()
+                if current_time - last_cancel_check >= cancel_check_interval:
+                    if not is_connected():
+                        # Stop signal detected during streaming
+                        yield Packet(
+                            placement=Placement(turn_index=last_turn_index + 1),
+                            obj=OverallStop(type="stop", stop_reason="user_cancelled"),
+                        )
+                        break
+                    last_cancel_check = current_time
     finally:
         # Wait for thread to complete on normal exit to propagate exceptions and ensure cleanup.
         # Skip waiting if user disconnected to exit quickly.

backend/onyx/chat/chat_utils.py

@@ -26,6 +26,7 @@ from onyx.context.search.models import RerankingDetails
 from onyx.context.search.models import RetrievalDetails
 from onyx.db.chat import create_chat_session
 from onyx.db.chat import get_chat_messages_by_session
+from onyx.db.chat import get_or_create_root_message
 from onyx.db.kg_config import get_kg_config_settings
 from onyx.db.kg_config import is_kg_config_settings_enabled_valid
 from onyx.db.llm import fetch_existing_doc_sets
@@ -37,6 +38,7 @@ from onyx.db.models import SearchDoc as DbSearchDoc
 from onyx.db.models import Tool
 from onyx.db.models import User
 from onyx.db.models import UserFile
+from onyx.db.projects import check_project_ownership
 from onyx.db.search_settings import get_current_search_settings
 from onyx.file_processing.extract_file_text import extract_file_text
 from onyx.file_store.file_store import get_default_file_store
@@ -51,6 +53,7 @@ from onyx.natural_language_processing.utils import BaseTokenizer
 from onyx.prompts.chat_prompts import ADDITIONAL_CONTEXT_PROMPT
 from onyx.prompts.chat_prompts import TOOL_CALL_RESPONSE_CROSS_MESSAGE
 from onyx.prompts.tool_prompts import TOOL_CALL_FAILURE_PROMPT
+from onyx.server.query_and_chat.models import ChatSessionCreationRequest
 from onyx.server.query_and_chat.models import CreateChatMessageRequest
 from onyx.server.query_and_chat.streaming_models import CitationInfo
 from onyx.tools.models import ToolCallKickoff
@@ -61,9 +64,45 @@ from onyx.utils.logger import setup_logger
 from onyx.utils.threadpool_concurrency import run_functions_tuples_in_parallel
 from onyx.utils.timing import log_function_time
 
+
 logger = setup_logger()
 
 
+def create_chat_session_from_request(
+    chat_session_request: ChatSessionCreationRequest,
+    user_id: UUID | None,
+    db_session: Session,
+) -> ChatSession:
+    """Create a chat session from a ChatSessionCreationRequest.
+
+    Includes project ownership validation when project_id is provided.
+
+    Args:
+        chat_session_request: The request containing persona_id, description, and project_id
+        user_id: The ID of the user creating the session (can be None for anonymous)
+        db_session: The database session
+
+    Returns:
+        The newly created ChatSession
+
+    Raises:
+        ValueError: If user lacks access to the specified project
+        Exception: If the persona is invalid
+    """
+    project_id = chat_session_request.project_id
+    if project_id:
+        if not check_project_ownership(project_id, user_id, db_session):
+            raise ValueError("User does not have access to project")
+
+    return create_chat_session(
+        db_session=db_session,
+        description=chat_session_request.description or "",
+        user_id=user_id,
+        persona_id=chat_session_request.persona_id,
+        project_id=chat_session_request.project_id,
+    )
+
+
 def prepare_chat_message_request(
     message_text: str,
     user: User | None,
@@ -77,6 +116,7 @@ def prepare_chat_message_request(
     skip_gen_ai_answer_generation: bool = False,
     llm_override: LLMOverride | None = None,
     allowed_tool_ids: list[int] | None = None,
+    forced_tool_ids: list[int] | None = None,
 ) -> CreateChatMessageRequest:
     # Typically used for one shot flows like SlackBot or non-chat API endpoint use cases
     new_chat_session = create_chat_session(
@@ -103,6 +143,7 @@ def prepare_chat_message_request(
         skip_gen_ai_answer_generation=skip_gen_ai_answer_generation,
         llm_override=llm_override,
         allowed_tool_ids=allowed_tool_ids,
+        forced_tool_ids=forced_tool_ids,
     )
 
 
@@ -164,13 +205,15 @@ def create_chat_history_chain(
     )
 
     if not all_chat_messages:
-        raise RuntimeError("No messages in Chat Session")
-
-    root_message = all_chat_messages[0]
-    if root_message.parent_message is not None:
-        raise RuntimeError(
-            "Invalid root message, unable to fetch valid chat message sequence"
+        root_message = get_or_create_root_message(
+            chat_session_id=chat_session_id, db_session=db_session
         )
+    else:
+        root_message = all_chat_messages[0]
+        if root_message.parent_message is not None:
+            raise RuntimeError(
+                "Invalid root message, unable to fetch valid chat message sequence"
+            )
 
     current_message: ChatMessage | None = root_message
     previous_message: ChatMessage | None = None
@@ -201,9 +244,6 @@ def create_chat_history_chain(
 
         previous_message = current_message
 
-    if not mainline_messages:
-        raise RuntimeError("Could not trace chat message history")
-
     return mainline_messages
 
 

backend/onyx/chat/llm_loop.py

@@ -99,7 +99,7 @@ def _build_project_file_citation_mapping(
 
 
 def construct_message_history(
-    system_prompt: ChatMessageSimple,
+    system_prompt: ChatMessageSimple | None,
     custom_agent_prompt: ChatMessageSimple | None,
     simple_chat_history: list[ChatMessageSimple],
     reminder_message: ChatMessageSimple | None,
@@ -114,7 +114,7 @@ def construct_message_history(
             )
 
     history_token_budget = available_tokens
-    history_token_budget -= system_prompt.token_count
+    history_token_budget -= system_prompt.token_count if system_prompt else 0
     history_token_budget -= (
         custom_agent_prompt.token_count if custom_agent_prompt else 0
     )
@@ -125,9 +125,12 @@ def construct_message_history(
     if history_token_budget < 0:
         raise ValueError("Not enough tokens available to construct message history")
 
+    if system_prompt:
+        system_prompt.should_cache = True
+
     # If no history, build minimal context
     if not simple_chat_history:
-        result = [system_prompt]
+        result = [system_prompt] if system_prompt else []
         if custom_agent_prompt:
             result.append(custom_agent_prompt)
         if project_files and project_files.project_file_texts:
@@ -199,6 +202,7 @@ def construct_message_history(
 
     for msg in reversed(history_before_last_user):
         if current_token_count + msg.token_count <= remaining_budget:
+            msg.should_cache = True
             truncated_history_before.insert(0, msg)
             current_token_count += msg.token_count
         else:
@@ -218,7 +222,7 @@ def construct_message_history(
     # Build the final message list according to README ordering:
     # [system], [history_before_last_user], [custom_agent], [project_files],
     # [last_user_message], [messages_after_last_user], [reminder]
-    result = [system_prompt]
+    result = [system_prompt] if system_prompt else []
 
     # 1. Add truncated history before last user message
     result.extend(truncated_history_before)
@@ -342,8 +346,13 @@ def run_llm_loop(
         has_called_search_tool: bool = False
         citation_mapping: dict[int, str] = {}  # Maps citation_num -> document_id/URL
 
+        default_base_system_prompt: str = get_default_base_system_prompt(db_session)
+        system_prompt = None
+        custom_agent_prompt_msg = None
+
         reasoning_cycles = 0
         for llm_cycle_count in range(MAX_LLM_CYCLES):
+            out_of_cycles = llm_cycle_count == MAX_LLM_CYCLES - 1
             if forced_tool_id:
                 # Needs to be just the single one because the "required" currently doesn't have a specified tool, just a binary
                 final_tools = [tool for tool in tools if tool.id == forced_tool_id]
@@ -351,7 +360,7 @@ def run_llm_loop(
                     raise ValueError(f"Tool {forced_tool_id} not found in tools")
                 tool_choice = ToolChoiceOptions.REQUIRED
                 forced_tool_id = None
-            elif llm_cycle_count == MAX_LLM_CYCLES - 1 or ran_image_gen:
+            elif out_of_cycles or ran_image_gen:
                 # Last cycle, no tools allowed, just answer!
                 tool_choice = ToolChoiceOptions.NONE
                 final_tools = []
@@ -370,35 +379,47 @@ def run_llm_loop(
                 )
                 custom_agent_prompt_msg = None
             else:
-                # System message and custom agent message are both included.
-                open_ai_formatting_enabled = model_needs_formatting_reenabled(
-                    llm.config.model_name
-                )
-
-                system_prompt_str = build_system_prompt(
-                    base_system_prompt=get_default_base_system_prompt(db_session),
-                    datetime_aware=persona.datetime_aware if persona else True,
-                    memories=memories,
-                    tools=tools,
-                    should_cite_documents=should_cite_documents
-                    or always_cite_documents,
-                    open_ai_formatting_enabled=open_ai_formatting_enabled,
-                )
-                system_prompt = ChatMessageSimple(
-                    message=system_prompt_str,
-                    token_count=token_counter(system_prompt_str),
-                    message_type=MessageType.SYSTEM,
-                )
-
-                custom_agent_prompt_msg = (
-                    ChatMessageSimple(
-                        message=custom_agent_prompt,
-                        token_count=token_counter(custom_agent_prompt),
-                        message_type=MessageType.USER,
+                # If it's an empty string, we assume the user does not want to include it as an empty System message
+                if default_base_system_prompt:
+                    open_ai_formatting_enabled = model_needs_formatting_reenabled(
+                        llm.config.model_name
                     )
-                    if custom_agent_prompt
-                    else None
-                )
+
+                    system_prompt_str = build_system_prompt(
+                        base_system_prompt=default_base_system_prompt,
+                        datetime_aware=persona.datetime_aware if persona else True,
+                        memories=memories,
+                        tools=tools,
+                        should_cite_documents=should_cite_documents
+                        or always_cite_documents,
+                        open_ai_formatting_enabled=open_ai_formatting_enabled,
+                    )
+                    system_prompt = ChatMessageSimple(
+                        message=system_prompt_str,
+                        token_count=token_counter(system_prompt_str),
+                        message_type=MessageType.SYSTEM,
+                    )
+                    custom_agent_prompt_msg = (
+                        ChatMessageSimple(
+                            message=custom_agent_prompt,
+                            token_count=token_counter(custom_agent_prompt),
+                            message_type=MessageType.USER,
+                        )
+                        if custom_agent_prompt
+                        else None
+                    )
+                else:
+                    # If there is a custom agent prompt, it replaces the system prompt when the default system prompt is empty
+                    system_prompt = (
+                        ChatMessageSimple(
+                            message=custom_agent_prompt,
+                            token_count=token_counter(custom_agent_prompt),
+                            message_type=MessageType.SYSTEM,
+                        )
+                        if custom_agent_prompt
+                        else None
+                    )
+                    custom_agent_prompt_msg = None
 
             reminder_message_text: str | None
             if ran_image_gen:
@@ -406,7 +427,7 @@ def run_llm_loop(
                 # This is to prevent it generating things like:
                 # [Cute Cat](attachment://a_cute_cat_sitting_playfully.png)
                 reminder_message_text = IMAGE_GEN_REMINDER
-            elif just_ran_web_search:
+            elif just_ran_web_search and not out_of_cycles:
                 reminder_message_text = OPEN_URL_REMINDER
             else:
                 # This is the default case, the LLM at this point may answer so it is important
@@ -417,6 +438,7 @@ def run_llm_loop(
                     ),
                     include_citation_reminder=should_cite_documents
                     or always_cite_documents,
+                    is_last_cycle=out_of_cycles,
                 )
 
             reminder_msg = (
@@ -491,6 +513,7 @@ def run_llm_loop(
                 user_info=None,  # TODO, this is part of memories right now, might want to separate it out
                 citation_mapping=citation_mapping,
                 next_citation_num=citation_processor.get_next_citation_number(),
+                max_concurrent_tools=None,
                 skip_search_query_expansion=has_called_search_tool,
             )
 

backend/onyx/chat/llm_step.py

@@ -1,4 +1,5 @@
 import json
+import time
 from collections.abc import Callable
 from collections.abc import Generator
 from collections.abc import Mapping
@@ -17,6 +18,7 @@ from onyx.context.search.models import SearchDoc
 from onyx.file_store.models import ChatFileType
 from onyx.llm.interfaces import LanguageModelInput
 from onyx.llm.interfaces import LLM
+from onyx.llm.interfaces import LLMConfig
 from onyx.llm.interfaces import LLMUserIdentity
 from onyx.llm.interfaces import ToolChoiceOptions
 from onyx.llm.model_response import Delta
@@ -31,6 +33,7 @@ from onyx.llm.models import TextContentPart
 from onyx.llm.models import ToolCall
 from onyx.llm.models import ToolMessage
 from onyx.llm.models import UserMessage
+from onyx.llm.prompt_cache.processor import process_with_prompt_cache
 from onyx.server.query_and_chat.placement import Placement
 from onyx.server.query_and_chat.streaming_models import AgentResponseDelta
 from onyx.server.query_and_chat.streaming_models import AgentResponseStart
@@ -46,7 +49,6 @@ from onyx.tracing.framework.create import generation_span
 from onyx.utils.b64 import get_image_type_from_bytes
 from onyx.utils.logger import setup_logger
 
-
 logger = setup_logger()
 
 
@@ -277,6 +279,7 @@ def _extract_tool_call_kickoffs(
 
 def translate_history_to_llm_format(
     history: list[ChatMessageSimple],
+    llm_config: LLMConfig,
 ) -> LanguageModelInput:
     """Convert a list of ChatMessageSimple to LanguageModelInput format.
 
@@ -284,8 +287,23 @@ def translate_history_to_llm_format(
     handling different message types and image files for multimodal support.
     """
     messages: list[ChatCompletionMessage] = []
+    last_cacheable_msg_idx = -1
+    all_previous_msgs_cacheable = True
+
+    for idx, msg in enumerate(history):
+        # if the message is being added to the history
+        if msg.message_type in [
+            MessageType.SYSTEM,
+            MessageType.USER,
+            MessageType.ASSISTANT,
+            MessageType.TOOL_CALL_RESPONSE,
+        ]:
+            all_previous_msgs_cacheable = (
+                all_previous_msgs_cacheable and msg.should_cache
+            )
+            if all_previous_msgs_cacheable:
+                last_cacheable_msg_idx = idx
 
-    for msg in history:
         if msg.message_type == MessageType.SYSTEM:
             system_msg = SystemMessage(
                 role="system",
@@ -395,7 +413,7 @@ def translate_history_to_llm_format(
             assistant_msg_with_tool = AssistantMessage(
                 role="assistant",
                 content=None,  # The tool call is parsed, doesn't need to be duplicated in the content
-                tool_calls=tool_calls if tool_calls else None,
+                tool_calls=tool_calls or None,
             )
             messages.append(assistant_msg_with_tool)
 
@@ -417,6 +435,18 @@ def translate_history_to_llm_format(
                 f"Unknown message type {msg.message_type} in history. Skipping message."
             )
 
+    # prompt caching: rely on should_cache in ChatMessageSimple to
+    # pick the split point for the cacheable prefix and suffix
+    if last_cacheable_msg_idx != -1:
+        processed_messages, _ = process_with_prompt_cache(
+            llm_config=llm_config,
+            cacheable_prefix=messages[: last_cacheable_msg_idx + 1],
+            suffix=messages[last_cacheable_msg_idx + 1 :],
+            continuation=False,
+        )
+        assert isinstance(processed_messages, list)  # for mypy
+        messages = processed_messages
+
     return messages
 
 
@@ -429,6 +459,10 @@ def _increment_turns(
         return turn_index, sub_turn_index + 1
 
 
+def _delta_has_action(delta: Delta) -> bool:
+    return bool(delta.content or delta.reasoning_content or delta.tool_calls)
+
+
 def run_llm_step_pkt_generator(
     history: list[ChatMessageSimple],
     tool_definitions: list[dict],
@@ -499,7 +533,7 @@ def run_llm_step_pkt_generator(
     tab_index = placement.tab_index
     sub_turn_index = placement.sub_turn_index
 
-    llm_msg_history = translate_history_to_llm_format(history)
+    llm_msg_history = translate_history_to_llm_format(history, llm.config)
     has_reasoned = 0
 
     # Uncomment the line below to log the entire message history to the console
@@ -526,6 +560,8 @@ def run_llm_step_pkt_generator(
         span_generation.span_data.input = cast(
             Sequence[Mapping[str, Any]], llm_msg_history
         )
+        stream_start_time = time.monotonic()
+        first_action_recorded = False
         for packet in llm.stream(
             prompt=llm_msg_history,
             tools=tool_definitions,
@@ -543,7 +579,13 @@ def run_llm_step_pkt_generator(
                     "cache_read_input_tokens": usage.cache_read_input_tokens,
                     "cache_creation_input_tokens": usage.cache_creation_input_tokens,
                 }
+                # Note: LLM cost tracking is now handled in multi_llm.py
             delta = packet.choice.delta
+            if not first_action_recorded and _delta_has_action(delta):
+                span_generation.span_data.time_to_first_action_seconds = (
+                    time.monotonic() - stream_start_time
+                )
+                first_action_recorded = True
 
             if custom_token_processor:
                 # The custom token processor can modify the deltas for specific custom logic
@@ -702,6 +744,15 @@ def run_llm_step_pkt_generator(
         # Flush custom token processor to get any final tool calls
         if custom_token_processor:
             flush_delta, processor_state = custom_token_processor(None, processor_state)
+            if (
+                not first_action_recorded
+                and flush_delta is not None
+                and _delta_has_action(flush_delta)
+            ):
+                span_generation.span_data.time_to_first_action_seconds = (
+                    time.monotonic() - stream_start_time
+                )
+                first_action_recorded = True
             if flush_delta and flush_delta.tool_calls:
                 for tool_call_delta in flush_delta.tool_calls:
                     _update_tool_call_with_delta(id_to_tool_call_map, tool_call_delta)

backend/onyx/chat/models.py

@@ -3,6 +3,7 @@ from collections.abc import Iterator
 from datetime import datetime
 from enum import Enum
 from typing import Any
+from uuid import UUID
 
 from pydantic import BaseModel
 from pydantic import Field
@@ -16,7 +17,9 @@ from onyx.context.search.models import SearchDoc
 from onyx.file_store.models import FileDescriptor
 from onyx.file_store.models import InMemoryChatFile
 from onyx.server.query_and_chat.streaming_models import CitationInfo
+from onyx.server.query_and_chat.streaming_models import GeneratedImage
 from onyx.server.query_and_chat.streaming_models import Packet
+from onyx.tools.models import SearchToolUsage
 from onyx.tools.models import ToolCallKickoff
 from onyx.tools.tool_implementations.custom.base_tool_types import ToolResultType
 
@@ -132,6 +135,13 @@ class ToolConfig(BaseModel):
     id: int
 
 
+class ProjectSearchConfig(BaseModel):
+    """Configuration for search tool availability in project context."""
+
+    search_usage: SearchToolUsage
+    disable_forced_tool: bool
+
+
 class PromptOverrideConfig(BaseModel):
     name: str
     description: str = ""
@@ -171,6 +181,10 @@ AnswerQuestionPossibleReturn = (
 )
 
 
+class CreateChatSessionID(BaseModel):
+    chat_session_id: UUID
+
+
 AnswerQuestionStreamReturn = Iterator[AnswerQuestionPossibleReturn]
 
 
@@ -181,12 +195,14 @@ class LLMMetricsContainer(BaseModel):
 
 StreamProcessor = Callable[[Iterator[str]], AnswerQuestionStreamReturn]
 
+
 AnswerStreamPart = (
     Packet
     | StreamStopInfo
     | MessageResponseIDInfo
     | StreamingError
     | UserKnowledgeFilePacket
+    | CreateChatSessionID
 )
 
 AnswerStream = Iterator[AnswerStreamPart]
@@ -204,6 +220,37 @@ class ChatBasicResponse(BaseModel):
     citation_info: list[CitationInfo]
 
 
+class ToolCallResponse(BaseModel):
+    """Tool call with full details for non-streaming response."""
+
+    tool_name: str
+    tool_arguments: dict[str, Any]
+    tool_result: str
+    search_docs: list[SearchDoc] | None = None
+    generated_images: list[GeneratedImage] | None = None
+    # Reasoning that led to the tool call
+    pre_reasoning: str | None = None
+
+
+class ChatFullResponse(BaseModel):
+    """Complete non-streaming response with all available data."""
+
+    # Core response fields
+    answer: str
+    answer_citationless: str
+    pre_answer_reasoning: str | None = None
+    tool_calls: list[ToolCallResponse] = []
+
+    # Documents & citations
+    top_documents: list[SearchDoc]
+    citation_info: list[CitationInfo]
+
+    # Metadata
+    message_id: int
+    chat_session_id: UUID | None = None
+    error_msg: str | None = None
+
+
 class ChatLoadedFile(InMemoryChatFile):
     content_text: str | None
     token_count: int
@@ -217,6 +264,12 @@ class ChatMessageSimple(BaseModel):
     image_files: list[ChatLoadedFile] | None = None
     # Only for TOOL_CALL_RESPONSE type messages
     tool_call_id: str | None = None
+    # The last message for which this is true
+    # AND is true for all previous messages
+    # (counting from the start of the history)
+    # represents the end of the cacheable prefix
+    # used for prompt caching
+    should_cache: bool = False
 
 
 class ProjectFileMetadata(BaseModel):
@@ -234,6 +287,8 @@ class ExtractedProjectFiles(BaseModel):
     total_token_count: int
     # Metadata for project files to enable citations
     project_file_metadata: list[ProjectFileMetadata]
+    # None if not a project
+    project_uncapped_token_count: int | None
 
 
 class LlmStepResult(BaseModel):

backend/onyx/chat/process_message.py

@@ -1,7 +1,10 @@
+"""
+IMPORTANT: familiarize yourself with the design concepts prior to contributing to this file.
+An overview can be found in the README.md file in this directory.
+"""
+
 import re
 import traceback
-from collections.abc import Callable
-from collections.abc import Iterator
 from uuid import UUID
 
 from sqlalchemy.orm import Session
@@ -10,6 +13,7 @@ from onyx.chat.chat_state import ChatStateContainer
 from onyx.chat.chat_state import run_chat_loop_with_state_containers
 from onyx.chat.chat_utils import convert_chat_history
 from onyx.chat.chat_utils import create_chat_history_chain
+from onyx.chat.chat_utils import create_chat_session_from_request
 from onyx.chat.chat_utils import get_custom_agent_prompt
 from onyx.chat.chat_utils import is_last_assistant_message_clarification
 from onyx.chat.chat_utils import load_all_chat_files
@@ -17,37 +21,36 @@ from onyx.chat.emitter import get_default_emitter
 from onyx.chat.llm_loop import run_llm_loop
 from onyx.chat.models import AnswerStream
 from onyx.chat.models import ChatBasicResponse
+from onyx.chat.models import ChatFullResponse
 from onyx.chat.models import ChatLoadedFile
+from onyx.chat.models import CreateChatSessionID
 from onyx.chat.models import ExtractedProjectFiles
 from onyx.chat.models import MessageResponseIDInfo
 from onyx.chat.models import ProjectFileMetadata
+from onyx.chat.models import ProjectSearchConfig
 from onyx.chat.models import StreamingError
+from onyx.chat.models import ToolCallResponse
 from onyx.chat.prompt_utils import calculate_reserved_tokens
 from onyx.chat.save_chat import save_chat_turn
 from onyx.chat.stop_signal_checker import is_connected as check_stop_signal
 from onyx.chat.stop_signal_checker import reset_cancel_status
-from onyx.configs.chat_configs import CHAT_TARGET_CHUNK_PERCENTAGE
-from onyx.configs.chat_configs import MAX_CHUNKS_FED_TO_CHAT
 from onyx.configs.constants import DEFAULT_PERSONA_ID
 from onyx.configs.constants import MessageType
 from onyx.configs.constants import MilestoneRecordType
+from onyx.context.search.enums import OptionalSearchSetting
 from onyx.context.search.models import CitationDocInfo
 from onyx.context.search.models import SearchDoc
 from onyx.db.chat import create_new_chat_message
-from onyx.db.chat import get_chat_message
 from onyx.db.chat import get_chat_session_by_id
 from onyx.db.chat import get_or_create_root_message
 from onyx.db.chat import reserve_message_id
-from onyx.db.engine.sql_engine import get_session_with_current_tenant
 from onyx.db.memory import get_memories
-from onyx.db.models import ChatMessage
 from onyx.db.models import User
 from onyx.db.projects import get_project_token_count
 from onyx.db.projects import get_user_files_from_project
 from onyx.db.tools import get_tools
 from onyx.deep_research.dr_loop import run_deep_research_llm_loop
 from onyx.file_store.models import ChatFileType
-from onyx.file_store.models import FileDescriptor
 from onyx.file_store.utils import load_in_memory_chat_files
 from onyx.file_store.utils import verify_user_files
 from onyx.llm.factory import get_llm_for_persona
@@ -57,23 +60,24 @@ from onyx.llm.interfaces import LLMUserIdentity
 from onyx.llm.utils import litellm_exception_to_error_msg
 from onyx.onyxbot.slack.models import SlackContext
 from onyx.redis.redis_pool import get_redis_client
+from onyx.server.query_and_chat.models import AUTO_PLACE_AFTER_LATEST_MESSAGE
 from onyx.server.query_and_chat.models import CreateChatMessageRequest
+from onyx.server.query_and_chat.models import SendMessageRequest
 from onyx.server.query_and_chat.streaming_models import AgentResponseDelta
 from onyx.server.query_and_chat.streaming_models import AgentResponseStart
 from onyx.server.query_and_chat.streaming_models import CitationInfo
 from onyx.server.query_and_chat.streaming_models import Packet
-from onyx.server.utils import get_json_line
+from onyx.server.usage_limits import check_llm_cost_limit_for_provider
 from onyx.tools.constants import SEARCH_TOOL_ID
 from onyx.tools.interface import Tool
+from onyx.tools.models import SearchToolUsage
 from onyx.tools.tool_constructor import construct_tools
 from onyx.tools.tool_constructor import CustomToolConfig
 from onyx.tools.tool_constructor import SearchToolConfig
-from onyx.tools.tool_constructor import SearchToolUsage
 from onyx.utils.logger import setup_logger
 from onyx.utils.long_term_log import LongTermLogger
 from onyx.utils.telemetry import mt_cloud_telemetry
 from onyx.utils.timing import log_function_time
-from onyx.utils.timing import log_generator_function_time
 from shared_configs.contextvars import get_current_tenant_id
 
 logger = setup_logger()
@@ -126,6 +130,7 @@ def _extract_project_file_texts_and_images(
             project_as_filter=False,
             total_token_count=0,
             project_file_metadata=[],
+            project_uncapped_token_count=None,
         )
 
     max_actual_tokens = (
@@ -211,159 +216,120 @@ def _extract_project_file_texts_and_images(
         project_as_filter=project_as_filter,
         total_token_count=total_token_count,
         project_file_metadata=project_file_metadata,
+        project_uncapped_token_count=project_tokens,
     )
 
 
 def _get_project_search_availability(
     project_id: int | None,
     persona_id: int | None,
-    has_project_file_texts: bool,
-    forced_tool_ids: list[int] | None,
+    loaded_project_files: bool,
+    project_has_files: bool,
+    forced_tool_id: int | None,
     search_tool_id: int | None,
-) -> SearchToolUsage:
+) -> ProjectSearchConfig:
     """Determine search tool availability based on project context.
 
-    Args:
-        project_id: The project ID if the user is in a project
-        persona_id: The persona ID to check if it's the default persona
-        has_project_file_texts: Whether project files are loaded in context
-        forced_tool_ids: List of forced tool IDs (may be mutated to remove search tool)
-        search_tool_id: The search tool ID to check against
+    Search is disabled when ALL of the following are true:
+    - User is in a project
+    - Using the default persona (not a custom agent)
+    - Project files are already loaded in context
 
-    Returns:
-        SearchToolUsage setting indicating how search should be used
+    When search is disabled and the user tried to force the search tool,
+    that forcing is also disabled.
+
+    Returns AUTO (follow persona config) in all other cases.
     """
-    # There are cases where the internal search tool should be disabled
-    # If the user is in a project, it should not use other sources / generic search
-    # If they are in a project but using a custom agent, it should use the agent setup
-    # (which means it can use search)
-    # However if in a project and there are more files than can fit in the context,
-    # it should use the search tool with the project filter on
-    # If no files are uploaded, search should remain enabled
-    search_usage_forcing_setting = SearchToolUsage.AUTO
-    if project_id:
-        if bool(persona_id is DEFAULT_PERSONA_ID and has_project_file_texts):
-            search_usage_forcing_setting = SearchToolUsage.DISABLED
-            # Remove search tool from forced_tool_ids if it's present
-            if forced_tool_ids and search_tool_id and search_tool_id in forced_tool_ids:
-                forced_tool_ids[:] = [
-                    tool_id for tool_id in forced_tool_ids if tool_id != search_tool_id
-                ]
-        elif forced_tool_ids and search_tool_id and search_tool_id in forced_tool_ids:
-            search_usage_forcing_setting = SearchToolUsage.ENABLED
-    return search_usage_forcing_setting
-
-
-def _initialize_chat_session(
-    message_text: str,
-    files: list[FileDescriptor],
-    token_counter: Callable[[str], int],
-    parent_id: int | None,
-    user_id: UUID | None,
-    chat_session_id: UUID,
-    db_session: Session,
-    use_existing_user_message: bool = False,
-) -> ChatMessage:
-    root_message = get_or_create_root_message(
-        chat_session_id=chat_session_id, db_session=db_session
-    )
-
-    if parent_id is None:
-        parent_message = root_message
-    else:
-        parent_message = get_chat_message(
-            chat_message_id=parent_id,
-            user_id=user_id,
-            db_session=db_session,
+    # Not in a project, this should have no impact on search tool availability
+    if not project_id:
+        return ProjectSearchConfig(
+            search_usage=SearchToolUsage.AUTO, disable_forced_tool=False
         )
 
-    # For seeding, the parent message points to the message that is supposed to be the last
-    # user message.
-    if use_existing_user_message:
-        if parent_message.parent_message is None:
-            raise RuntimeError("No parent message found for seeding")
-        if parent_message.message_type != MessageType.USER:
-            raise RuntimeError(
-                "Parent message is not a user message, needed for seeded flow."
-            )
-        message_text = parent_message.message
-        token_count = parent_message.token_count
-        parent_message = parent_message.parent_message
-    else:
-        token_count = token_counter(message_text)
+    # Custom persona in project - let persona config decide
+    # Even if there are no files in the project, it's still guided by the persona config.
+    if persona_id != DEFAULT_PERSONA_ID:
+        return ProjectSearchConfig(
+            search_usage=SearchToolUsage.AUTO, disable_forced_tool=False
+        )
 
-    # Flushed for ID but not committed yet
-    user_message = create_new_chat_message(
-        chat_session_id=chat_session_id,
-        parent_message=parent_message,
-        message=message_text,
-        token_count=token_count,
-        message_type=MessageType.USER,
-        files=files,
-        db_session=db_session,
-        commit=False,
+    # If in a project with the default persona and the files have been already loaded into the context or
+    # there are no files in the project, disable search as there is nothing to search for.
+    if loaded_project_files or not project_has_files:
+        user_forced_search = (
+            forced_tool_id is not None
+            and search_tool_id is not None
+            and forced_tool_id == search_tool_id
+        )
+        return ProjectSearchConfig(
+            search_usage=SearchToolUsage.DISABLED,
+            disable_forced_tool=user_forced_search,
+        )
+
+    # Default persona in a project with files, but also the files have not been loaded into the context already.
+    return ProjectSearchConfig(
+        search_usage=SearchToolUsage.ENABLED, disable_forced_tool=False
     )
-    return user_message
 
 
-def stream_chat_message_objects(
-    new_msg_req: CreateChatMessageRequest,
+def handle_stream_message_objects(
+    new_msg_req: SendMessageRequest,
     user: User | None,
     db_session: Session,
-    # Needed to translate persona num_chunks to tokens to the LLM
-    default_num_chunks: float = MAX_CHUNKS_FED_TO_CHAT,
-    # For flow with search, don't include as many chunks as possible since we need to leave space
-    # for the chat history, for smaller models, we likely won't get MAX_CHUNKS_FED_TO_CHAT chunks
-    max_document_percentage: float = CHAT_TARGET_CHUNK_PERCENTAGE,
     # if specified, uses the last user message and does not create a new user message based
     # on the `new_msg_req.message`. Currently, requires a state where the last message is a
     litellm_additional_headers: dict[str, str] | None = None,
     custom_tool_additional_headers: dict[str, str] | None = None,
-    is_connected: Callable[[], bool] | None = None,
-    enforce_chat_session_id_for_search_docs: bool = True,
     bypass_acl: bool = False,
     # Additional context that should be included in the chat history, for example:
     # Slack threads where the conversation cannot be represented by a chain of User/Assistant
-    # messages.
+    # messages. Both of the below are used for Slack
     # NOTE: is not stored in the database, only passed in to the LLM as context
     additional_context: str | None = None,
     # Slack context for federated Slack search
     slack_context: SlackContext | None = None,
+    # Optional external state container for non-streaming access to accumulated state
+    external_state_container: ChatStateContainer | None = None,
 ) -> AnswerStream:
     tenant_id = get_current_tenant_id()
-    use_existing_user_message = new_msg_req.use_existing_user_message
 
     llm: LLM | None = None
 
+    user_id = user.id if user is not None else None
+    llm_user_identifier = (
+        user.email
+        if user is not None and getattr(user, "email", None)
+        else (str(user_id) if user_id else "anonymous_user")
+    )
     try:
-        user_id = user.id if user is not None else None
-        llm_user_identifier = (
-            user.email
-            if user is not None and getattr(user, "email", None)
-            else (str(user_id) if user_id else "anonymous_user")
-        )
+        if not new_msg_req.chat_session_id:
+            if not new_msg_req.chat_session_info:
+                raise RuntimeError(
+                    "Must specify a chat session id or chat session info"
+                )
+            chat_session = create_chat_session_from_request(
+                chat_session_request=new_msg_req.chat_session_info,
+                user_id=user_id,
+                db_session=db_session,
+            )
+            yield CreateChatSessionID(chat_session_id=chat_session.id)
+        else:
+            chat_session = get_chat_session_by_id(
+                chat_session_id=new_msg_req.chat_session_id,
+                user_id=user_id,
+                db_session=db_session,
+            )
 
-        chat_session = get_chat_session_by_id(
-            chat_session_id=new_msg_req.chat_session_id,
-            user_id=user_id,
-            db_session=db_session,
-        )
         persona = chat_session.persona
 
         message_text = new_msg_req.message
-        chat_session_id = new_msg_req.chat_session_id
         user_identity = LLMUserIdentity(
-            user_id=llm_user_identifier, session_id=str(chat_session_id)
+            user_id=llm_user_identifier, session_id=str(chat_session.id)
         )
-        parent_id = new_msg_req.parent_message_id
-        reference_doc_ids = new_msg_req.search_doc_ids
-        retrieval_options = new_msg_req.retrieval_options
-        new_msg_req.alternate_assistant_id
-        user_selected_filters = retrieval_options.filters if retrieval_options else None
 
         # permanent "log" store, used primarily for debugging
         long_term_logger = LongTermLogger(
-            metadata={"user_id": str(user_id), "chat_session_id": str(chat_session_id)}
+            metadata={"user_id": str(user_id), "chat_session_id": str(chat_session.id)}
         )
 
         # Milestone tracking, most devs using the API don't need to understand this
@@ -373,11 +339,6 @@ def stream_chat_message_objects(
             event=MilestoneRecordType.MULTIPLE_ASSISTANTS,
         )
 
-        if reference_doc_ids is None and retrieval_options is None:
-            raise RuntimeError(
-                "Must specify a set of documents for chat or specify search options"
-            )
-
         llm = get_llm_for_persona(
             persona=persona,
             user=user,
@@ -387,6 +348,14 @@ def stream_chat_message_objects(
         )
         token_counter = get_llm_token_counter(llm)
 
+        # Check LLM cost limits before using the LLM (only for Onyx-managed keys)
+
+        check_llm_cost_limit_for_provider(
+            db_session=db_session,
+            tenant_id=tenant_id,
+            llm_provider_api_key=llm.config.api_key,
+        )
+
         # Verify that the user specified files actually belong to the user
         verify_user_files(
             user_files=new_msg_req.file_descriptors,
@@ -395,35 +364,58 @@ def stream_chat_message_objects(
             project_id=chat_session.project_id,
         )
 
-        # Makes sure that the chat session has the right message nodes
-        # and that the latest user message is created (not yet committed)
-        user_message = _initialize_chat_session(
-            message_text=message_text,
-            files=new_msg_req.file_descriptors,
-            token_counter=token_counter,
-            parent_id=parent_id,
-            user_id=user_id,
-            chat_session_id=chat_session_id,
-            db_session=db_session,
-            use_existing_user_message=use_existing_user_message,
-        )
-
         # re-create linear history of messages
         chat_history = create_chat_history_chain(
-            chat_session_id=chat_session_id, db_session=db_session
+            chat_session_id=chat_session.id, db_session=db_session
         )
 
-        last_chat_message = chat_history[-1]
+        # Determine the parent message based on the request:
+        # - -1: auto-place after latest message in chain
+        # - None: regeneration from root (first message)
+        # - positive int: place after that specific parent message
+        root_message = get_or_create_root_message(
+            chat_session_id=chat_session.id, db_session=db_session
+        )
 
-        if last_chat_message.id != user_message.id:
-            db_session.rollback()
-            raise RuntimeError(
-                "The new message was not on the mainline. "
-                "Chat message history tree is not correctly built."
+        if new_msg_req.parent_message_id == AUTO_PLACE_AFTER_LATEST_MESSAGE:
+            # Auto-place after the latest message in the chain
+            parent_message = chat_history[-1] if chat_history else root_message
+        elif new_msg_req.parent_message_id is None:
+            # None = regeneration from root
+            parent_message = root_message
+            # Truncate history since we're starting from root
+            chat_history = []
+        else:
+            # Specific parent message ID provided, find parent in chat_history
+            parent_message = None
+            for i in range(len(chat_history) - 1, -1, -1):
+                if chat_history[i].id == new_msg_req.parent_message_id:
+                    parent_message = chat_history[i]
+                    # Truncate history to only include messages up to and including parent
+                    chat_history = chat_history[: i + 1]
+                    break
+
+        if parent_message is None:
+            raise ValueError(
+                "The new message sent is not on the latest mainline of messages"
             )
 
-        # At this point we can save the user message as it's validated and final
-        db_session.commit()
+        # If the parent message is a user message, it's a regeneration and we use the existing user message.
+        if parent_message.message_type == MessageType.USER:
+            user_message = parent_message
+        else:
+            user_message = create_new_chat_message(
+                chat_session_id=chat_session.id,
+                parent_message=parent_message,
+                message=message_text,
+                token_count=token_counter(message_text),
+                message_type=MessageType.USER,
+                files=new_msg_req.file_descriptors,
+                db_session=db_session,
+                commit=True,
+            )
+
+            chat_history.append(user_message)
 
         memories = get_memories(user, db_session)
 
@@ -433,7 +425,7 @@ def stream_chat_message_objects(
             db_session=db_session,
             persona_system_prompt=custom_agent_prompt or "",
             token_counter=token_counter,
-            files=last_chat_message.files,
+            files=new_msg_req.file_descriptors,
             memories=memories,
         )
 
@@ -455,15 +447,20 @@ def stream_chat_message_objects(
             None,
         )
 
-        # This may also mutate the new_msg_req.forced_tool_ids
-        # This logic is specifically for projects
-        search_usage_forcing_setting = _get_project_search_availability(
+        # Determine if search should be disabled for this project context
+        forced_tool_id = new_msg_req.forced_tool_id
+        project_search_config = _get_project_search_availability(
             project_id=chat_session.project_id,
             persona_id=persona.id,
-            has_project_file_texts=bool(extracted_project_files.project_file_texts),
-            forced_tool_ids=new_msg_req.forced_tool_ids,
+            loaded_project_files=bool(extracted_project_files.project_file_texts),
+            project_has_files=bool(
+                extracted_project_files.project_uncapped_token_count
+            ),
+            forced_tool_id=new_msg_req.forced_tool_id,
             search_tool_id=search_tool_id,
         )
+        if project_search_config.disable_forced_tool:
+            forced_tool_id = None
 
         emitter = get_default_emitter()
 
@@ -475,7 +472,7 @@ def stream_chat_message_objects(
             user=user,
             llm=llm,
             search_tool_config=SearchToolConfig(
-                user_selected_filters=user_selected_filters,
+                user_selected_filters=new_msg_req.internal_search_filters,
                 project_id=(
                     chat_session.project_id
                     if extracted_project_files.project_as_filter
@@ -485,17 +482,20 @@ def stream_chat_message_objects(
                 slack_context=slack_context,
             ),
             custom_tool_config=CustomToolConfig(
-                chat_session_id=chat_session_id,
+                chat_session_id=chat_session.id,
                 message_id=user_message.id if user_message else None,
                 additional_headers=custom_tool_additional_headers,
             ),
             allowed_tool_ids=new_msg_req.allowed_tool_ids,
-            search_usage_forcing_setting=search_usage_forcing_setting,
+            search_usage_forcing_setting=project_search_config.search_usage,
         )
         tools: list[Tool] = []
         for tool_list in tool_dict.values():
             tools.extend(tool_list)
 
+        if forced_tool_id and forced_tool_id not in [tool.id for tool in tools]:
+            raise ValueError(f"Forced tool {forced_tool_id} not found in tools")
+
         # TODO Once summarization is done, we don't need to load all the files from the beginning anymore.
         # load all files needed for this chat chain in memory
         files = load_all_chat_files(chat_history, db_session)
@@ -505,7 +505,7 @@ def stream_chat_message_objects(
         # Reserve a message id for the assistant response for frontend to track packets
         assistant_response = reserve_message_id(
             db_session=db_session,
-            chat_session_id=chat_session_id,
+            chat_session_id=chat_session.id,
             parent_message=user_message.id,
             message_type=MessageType.ASSISTANT,
         )
@@ -529,15 +529,16 @@ def stream_chat_message_objects(
         redis_client = get_redis_client()
 
         reset_cancel_status(
-            chat_session_id,
+            chat_session.id,
             redis_client,
         )
 
         def check_is_connected() -> bool:
-            return check_stop_signal(chat_session_id, redis_client)
+            return check_stop_signal(chat_session.id, redis_client)
 
-        # Create state container for accumulating partial results
-        state_container = ChatStateContainer()
+        # Use external state container if provided, otherwise create internal one
+        # External container allows non-streaming callers to access accumulated state
+        state_container = external_state_container or ChatStateContainer()
 
         # Run the LLM loop with explicit wrapper for stop signal handling
         # The wrapper runs run_llm_loop in a background thread and polls every 300ms
@@ -565,7 +566,7 @@ def stream_chat_message_objects(
                 db_session=db_session,
                 skip_clarification=skip_clarification,
                 user_identity=user_identity,
-                chat_session_id=str(chat_session_id),
+                chat_session_id=str(chat_session.id),
             )
         else:
             yield from run_chat_loop_with_state_containers(
@@ -582,19 +583,15 @@ def stream_chat_message_objects(
                 llm=llm,
                 token_counter=token_counter,
                 db_session=db_session,
-                forced_tool_id=(
-                    new_msg_req.forced_tool_ids[0]
-                    if new_msg_req.forced_tool_ids
-                    else None
-                ),
+                forced_tool_id=forced_tool_id,
                 user_identity=user_identity,
-                chat_session_id=str(chat_session_id),
+                chat_session_id=str(chat_session.id),
             )
 
         # Determine if stopped by user
         completed_normally = check_is_connected()
         if not completed_normally:
-            logger.debug(f"Chat session {chat_session_id} stopped by user")
+            logger.debug(f"Chat session {chat_session.id} stopped by user")
 
         # Build final answer based on completion status
         if completed_normally:
@@ -696,23 +693,63 @@ def stream_chat_message_objects(
         return
 
 
-@log_generator_function_time()
-def stream_chat_message(
+def stream_chat_message_objects(
     new_msg_req: CreateChatMessageRequest,
     user: User | None,
+    db_session: Session,
+    # if specified, uses the last user message and does not create a new user message based
+    # on the `new_msg_req.message`. Currently, requires a state where the last message is a
     litellm_additional_headers: dict[str, str] | None = None,
     custom_tool_additional_headers: dict[str, str] | None = None,
-) -> Iterator[str]:
-    with get_session_with_current_tenant() as db_session:
-        objects = stream_chat_message_objects(
-            new_msg_req=new_msg_req,
-            user=user,
-            db_session=db_session,
-            litellm_additional_headers=litellm_additional_headers,
-            custom_tool_additional_headers=custom_tool_additional_headers,
+    bypass_acl: bool = False,
+    # Additional context that should be included in the chat history, for example:
+    # Slack threads where the conversation cannot be represented by a chain of User/Assistant
+    # messages. Both of the below are used for Slack
+    # NOTE: is not stored in the database, only passed in to the LLM as context
+    additional_context: str | None = None,
+    # Slack context for federated Slack search
+    slack_context: SlackContext | None = None,
+) -> AnswerStream:
+    forced_tool_id = (
+        new_msg_req.forced_tool_ids[0] if new_msg_req.forced_tool_ids else None
+    )
+    if (
+        new_msg_req.retrieval_options
+        and new_msg_req.retrieval_options.run_search == OptionalSearchSetting.ALWAYS
+    ):
+        all_tools = get_tools(db_session)
+
+        search_tool_id = next(
+            (tool.id for tool in all_tools if tool.in_code_tool_id == SEARCH_TOOL_ID),
+            None,
         )
-        for obj in objects:
-            yield get_json_line(obj.model_dump())
+        forced_tool_id = search_tool_id
+
+    translated_new_msg_req = SendMessageRequest(
+        message=new_msg_req.message,
+        llm_override=new_msg_req.llm_override,
+        allowed_tool_ids=new_msg_req.allowed_tool_ids,
+        forced_tool_id=forced_tool_id,
+        file_descriptors=new_msg_req.file_descriptors,
+        internal_search_filters=(
+            new_msg_req.retrieval_options.filters
+            if new_msg_req.retrieval_options
+            else None
+        ),
+        deep_research=new_msg_req.deep_research,
+        parent_message_id=new_msg_req.parent_message_id,
+        chat_session_id=new_msg_req.chat_session_id,
+    )
+    return handle_stream_message_objects(
+        new_msg_req=translated_new_msg_req,
+        user=user,
+        db_session=db_session,
+        litellm_additional_headers=litellm_additional_headers,
+        custom_tool_additional_headers=custom_tool_additional_headers,
+        bypass_acl=bypass_acl,
+        additional_context=additional_context,
+        slack_context=slack_context,
+    )
 
 
 def remove_answer_citations(answer: str) -> str:
@@ -767,3 +804,83 @@ def gather_stream(
         error_msg=error_msg,
         top_documents=top_documents,
     )
+
+
+@log_function_time()
+def gather_stream_full(
+    packets: AnswerStream,
+    state_container: ChatStateContainer,
+) -> ChatFullResponse:
+    """
+    Aggregate streaming packets and state container into a complete ChatFullResponse.
+
+    This function consumes all packets from the stream and combines them with
+    the accumulated state from the ChatStateContainer to build a complete response
+    including answer, reasoning, citations, and tool calls.
+
+    Args:
+        packets: The stream of packets from handle_stream_message_objects
+        state_container: The state container that accumulates tool calls, reasoning, etc.
+
+    Returns:
+        ChatFullResponse with all available data
+    """
+    answer: str | None = None
+    citations: list[CitationInfo] = []
+    error_msg: str | None = None
+    message_id: int | None = None
+    top_documents: list[SearchDoc] = []
+    chat_session_id: UUID | None = None
+
+    for packet in packets:
+        if isinstance(packet, Packet):
+            if isinstance(packet.obj, AgentResponseStart):
+                if packet.obj.final_documents:
+                    top_documents = packet.obj.final_documents
+            elif isinstance(packet.obj, AgentResponseDelta):
+                if answer is None:
+                    answer = ""
+                if packet.obj.content:
+                    answer += packet.obj.content
+            elif isinstance(packet.obj, CitationInfo):
+                citations.append(packet.obj)
+        elif isinstance(packet, StreamingError):
+            error_msg = packet.error
+        elif isinstance(packet, MessageResponseIDInfo):
+            message_id = packet.reserved_assistant_message_id
+        elif isinstance(packet, CreateChatSessionID):
+            chat_session_id = packet.chat_session_id
+
+    if message_id is None:
+        raise ValueError("Message ID is required")
+
+    # Use state_container for complete answer (handles edge cases gracefully)
+    final_answer = state_container.get_answer_tokens() or answer or ""
+
+    # Get reasoning from state container (None when model doesn't produce reasoning)
+    reasoning = state_container.get_reasoning_tokens()
+
+    # Convert ToolCallInfo list to ToolCallResponse list
+    tool_call_responses = [
+        ToolCallResponse(
+            tool_name=tc.tool_name,
+            tool_arguments=tc.tool_call_arguments,
+            tool_result=tc.tool_call_response,
+            search_docs=tc.search_docs,
+            generated_images=tc.generated_images,
+            pre_reasoning=tc.reasoning_tokens,
+        )
+        for tc in state_container.get_tool_calls()
+    ]
+
+    return ChatFullResponse(
+        answer=final_answer,
+        answer_citationless=remove_answer_citations(final_answer),
+        pre_answer_reasoning=reasoning,
+        tool_calls=tool_call_responses,
+        top_documents=top_documents,
+        citation_info=citations,
+        message_id=message_id,
+        chat_session_id=chat_session_id,
+        error_msg=error_msg,
+    )

backend/onyx/chat/prompt_utils.py

@@ -10,6 +10,7 @@ from onyx.file_store.models import FileDescriptor
 from onyx.prompts.chat_prompts import CITATION_REMINDER
 from onyx.prompts.chat_prompts import CODE_BLOCK_MARKDOWN
 from onyx.prompts.chat_prompts import DEFAULT_SYSTEM_PROMPT
+from onyx.prompts.chat_prompts import LAST_CYCLE_CITATION_REMINDER
 from onyx.prompts.chat_prompts import REQUIRE_CITATION_GUIDANCE
 from onyx.prompts.chat_prompts import USER_INFO_HEADER
 from onyx.prompts.prompt_utils import get_company_context
@@ -22,6 +23,7 @@ from onyx.prompts.tool_prompts import PYTHON_TOOL_GUIDANCE
 from onyx.prompts.tool_prompts import TOOL_DESCRIPTION_SEARCH_GUIDANCE
 from onyx.prompts.tool_prompts import TOOL_SECTION_HEADER
 from onyx.prompts.tool_prompts import WEB_SEARCH_GUIDANCE
+from onyx.prompts.tool_prompts import WEB_SEARCH_SITE_DISABLED_GUIDANCE
 from onyx.tools.interface import Tool
 from onyx.tools.tool_implementations.images.image_generation_tool import (
     ImageGenerationTool,
@@ -37,7 +39,7 @@ def get_default_base_system_prompt(db_session: Session) -> str:
     default_persona = get_default_behavior_persona(db_session)
     return (
         default_persona.system_prompt
-        if default_persona and default_persona.system_prompt
+        if default_persona and default_persona.system_prompt is not None
         else DEFAULT_SYSTEM_PROMPT
     )
 
@@ -115,8 +117,11 @@ def calculate_reserved_tokens(
 def build_reminder_message(
     reminder_text: str | None,
     include_citation_reminder: bool,
+    is_last_cycle: bool,
 ) -> str | None:
     reminder = reminder_text.strip() if reminder_text else ""
+    if is_last_cycle:
+        reminder += "\n\n" + LAST_CYCLE_CITATION_REMINDER
     if include_citation_reminder:
         reminder += "\n\n" + CITATION_REMINDER
     reminder = reminder.strip()
@@ -169,7 +174,9 @@ def build_system_prompt(
             TOOL_SECTION_HEADER
             + TOOL_DESCRIPTION_SEARCH_GUIDANCE
             + INTERNAL_SEARCH_GUIDANCE
-            + WEB_SEARCH_GUIDANCE
+            + WEB_SEARCH_GUIDANCE.format(
+                site_colon_disabled=WEB_SEARCH_SITE_DISABLED_GUIDANCE
+            )
             + OPEN_URLS_GUIDANCE
             + GENERATE_IMAGE_GUIDANCE
             + PYTHON_TOOL_GUIDANCE
@@ -195,7 +202,16 @@ def build_system_prompt(
             system_prompt += INTERNAL_SEARCH_GUIDANCE
 
         if has_web_search or include_all_guidance:
-            system_prompt += WEB_SEARCH_GUIDANCE
+            site_disabled_guidance = ""
+            if has_web_search:
+                web_search_tool = next(
+                    (t for t in tools if isinstance(t, WebSearchTool)), None
+                )
+                if web_search_tool and not web_search_tool.supports_site_filter:
+                    site_disabled_guidance = WEB_SEARCH_SITE_DISABLED_GUIDANCE
+            system_prompt += WEB_SEARCH_GUIDANCE.format(
+                site_colon_disabled=site_disabled_guidance
+            )
 
         if has_open_urls or include_all_guidance:
             system_prompt += OPEN_URLS_GUIDANCE

backend/onyx/chat/save_chat.py

@@ -117,22 +117,30 @@ def _create_and_link_tool_calls(
         tool_call_map[tool_call_obj.tool_call_id] = tool_call_obj.id
 
     # Update parent_tool_call_id for all tool calls
+    # Filter out orphaned children (whose parents don't exist) - this can happen
+    # when generation is stopped mid-execution and parent tool calls were cancelled
+    valid_tool_calls: list[ToolCall] = []
     for tool_call_obj in tool_call_objects:
         tool_call_info = tool_call_info_map[tool_call_obj.tool_call_id]
         if tool_call_info.parent_tool_call_id is not None:
             parent_id = tool_call_map.get(tool_call_info.parent_tool_call_id)
             if parent_id is not None:
                 tool_call_obj.parent_tool_call_id = parent_id
+                valid_tool_calls.append(tool_call_obj)
             else:
-                # This would cause chat sessions to fail if this function is miscalled with
-                # tool calls that have bad parent pointers but this falls under "fail loudly"
-                raise ValueError(
-                    f"Parent tool call with tool_call_id '{tool_call_info.parent_tool_call_id}' "
-                    f"not found for tool call '{tool_call_obj.tool_call_id}'"
+                # Parent doesn't exist (likely cancelled) - skip this orphaned child
+                logger.warning(
+                    f"Skipping tool call '{tool_call_obj.tool_call_id}' with missing parent "
+                    f"'{tool_call_info.parent_tool_call_id}' (likely cancelled during execution)"
                 )
+                # Remove from DB session to prevent saving
+                db_session.delete(tool_call_obj)
+        else:
+            # Top-level tool call (no parent)
+            valid_tool_calls.append(tool_call_obj)
 
-    # Link SearchDocs to ToolCalls
-    for tool_call_obj in tool_call_objects:
+    # Link SearchDocs only to valid ToolCalls
+    for tool_call_obj in valid_tool_calls:
         search_doc_ids = tool_call_to_search_doc_ids.get(tool_call_obj.tool_call_id, [])
         if search_doc_ids:
             add_search_docs_to_tool_call(

backend/onyx/chat/stop_signal_checker.py

@@ -2,12 +2,23 @@ from uuid import UUID
 
 from redis.client import Redis
 
-from shared_configs.contextvars import get_current_tenant_id
-
 # Redis key prefixes for chat session stop signals
 PREFIX = "chatsessionstop"
 FENCE_PREFIX = f"{PREFIX}_fence"
-FENCE_TTL = 24 * 60 * 60  # 24 hours - defensive TTL to prevent memory leaks
+FENCE_TTL = 10 * 60  # 10 minutes - defensive TTL to prevent memory leaks
+
+
+def _get_fence_key(chat_session_id: UUID) -> str:
+    """
+    Generate the Redis key for a chat session stop signal fence.
+
+    Args:
+        chat_session_id: The UUID of the chat session
+
+    Returns:
+        The fence key string (tenant_id is automatically added by the Redis client)
+    """
+    return f"{FENCE_PREFIX}_{chat_session_id}"
 
 
 def set_fence(chat_session_id: UUID, redis_client: Redis, value: bool) -> None:
@@ -16,11 +27,10 @@ def set_fence(chat_session_id: UUID, redis_client: Redis, value: bool) -> None:
 
     Args:
         chat_session_id: The UUID of the chat session
-        redis_client: Redis client to use
+        redis_client: Redis client to use (tenant-aware client that auto-prefixes keys)
         value: True to set the fence (stop signal), False to clear it
     """
-    tenant_id = get_current_tenant_id()
-    fence_key = f"{FENCE_PREFIX}_{tenant_id}_{chat_session_id}"
+    fence_key = _get_fence_key(chat_session_id)
     if not value:
         redis_client.delete(fence_key)
         return
@@ -34,13 +44,12 @@ def is_connected(chat_session_id: UUID, redis_client: Redis) -> bool:
 
     Args:
         chat_session_id: The UUID of the chat session to check
-        redis_client: Redis client to use for checking the stop signal
+        redis_client: Redis client to use for checking the stop signal (tenant-aware client that auto-prefixes keys)
 
     Returns:
         True if the session should continue, False if it should stop
     """
-    tenant_id = get_current_tenant_id()
-    fence_key = f"{FENCE_PREFIX}_{tenant_id}_{chat_session_id}"
+    fence_key = _get_fence_key(chat_session_id)
     return not bool(redis_client.exists(fence_key))
 
 
@@ -50,8 +59,7 @@ def reset_cancel_status(chat_session_id: UUID, redis_client: Redis) -> None:
 
     Args:
         chat_session_id: The UUID of the chat session
-        redis_client: Redis client to use
+        redis_client: Redis client to use (tenant-aware client that auto-prefixes keys)
     """
-    tenant_id = get_current_tenant_id()
-    fence_key = f"{FENCE_PREFIX}_{tenant_id}_{chat_session_id}"
+    fence_key = _get_fence_key(chat_session_id)
     redis_client.delete(fence_key)

backend/onyx/configs/app_configs.py

@@ -7,10 +7,8 @@ from typing import cast
 
 from onyx.auth.schemas import AuthBackend
 from onyx.configs.constants import AuthType
-from onyx.configs.constants import DocumentIndexType
 from onyx.configs.constants import QueryHistoryType
 from onyx.file_processing.enums import HtmlBasedConnectorTransformLinksStrategy
-from onyx.prompts.image_analysis import DEFAULT_IMAGE_ANALYSIS_SYSTEM_PROMPT
 from onyx.prompts.image_analysis import DEFAULT_IMAGE_SUMMARIZATION_SYSTEM_PROMPT
 from onyx.prompts.image_analysis import DEFAULT_IMAGE_SUMMARIZATION_USER_PROMPT
 
@@ -122,6 +120,14 @@ VALID_EMAIL_DOMAINS = (
     if _VALID_EMAIL_DOMAINS_STR
     else []
 )
+
+# Disposable email blocking - blocks temporary/throwaway email addresses
+# Set to empty string to disable disposable email blocking
+DISPOSABLE_EMAIL_DOMAINS_URL = os.environ.get(
+    "DISPOSABLE_EMAIL_DOMAINS_URL",
+    "https://disposable.github.io/disposable-email-domains/domains.json",
+)
+
 # OAuth Login Flow
 # Used for both Google OAuth2 and OIDC flows
 OAUTH_CLIENT_ID = (
@@ -188,10 +194,16 @@ TRACK_EXTERNAL_IDP_EXPIRY = (
 # DB Configs
 #####
 DOCUMENT_INDEX_NAME = "danswer_index"
-# Vespa is now the default document index store for both keyword and vector
-DOCUMENT_INDEX_TYPE = os.environ.get(
-    "DOCUMENT_INDEX_TYPE", DocumentIndexType.COMBINED.value
+
+OPENSEARCH_HOST = os.environ.get("OPENSEARCH_HOST") or "localhost"
+OPENSEARCH_REST_API_PORT = int(os.environ.get("OPENSEARCH_REST_API_PORT") or 9200)
+OPENSEARCH_ADMIN_USERNAME = os.environ.get("OPENSEARCH_ADMIN_USERNAME", "admin")
+OPENSEARCH_ADMIN_PASSWORD = os.environ.get("OPENSEARCH_ADMIN_PASSWORD", "")
+
+ENABLE_OPENSEARCH_FOR_ONYX = (
+    os.environ.get("ENABLE_OPENSEARCH_FOR_ONYX", "").lower() == "true"
 )
+
 VESPA_HOST = os.environ.get("VESPA_HOST") or "localhost"
 # NOTE: this is used if and only if the vespa config server is accessible via a
 # different host than the main vespa application
@@ -203,10 +215,6 @@ VESPA_NUM_ATTEMPTS_ON_STARTUP = int(os.environ.get("NUM_RETRIES_ON_STARTUP") or
 
 VESPA_CLOUD_URL = os.environ.get("VESPA_CLOUD_URL", "")
 
-# The default below is for dockerized deployment
-VESPA_DEPLOYMENT_ZIP = (
-    os.environ.get("VESPA_DEPLOYMENT_ZIP") or "/app/onyx/vespa-app.zip"
-)
 VESPA_CLOUD_CERT_PATH = os.environ.get("VESPA_CLOUD_CERT_PATH")
 VESPA_CLOUD_KEY_PATH = os.environ.get("VESPA_CLOUD_KEY_PATH")
 
@@ -315,63 +323,64 @@ CELERY_RESULT_EXPIRES = int(os.environ.get("CELERY_RESULT_EXPIRES", 86400))  # s
 
 # https://docs.celeryq.dev/en/stable/userguide/configuration.html#broker-pool-limit
 # Setting to None may help when there is a proxy in the way closing idle connections
-CELERY_BROKER_POOL_LIMIT_DEFAULT = 10
+_CELERY_BROKER_POOL_LIMIT_DEFAULT = 10
 try:
     CELERY_BROKER_POOL_LIMIT = int(
-        os.environ.get("CELERY_BROKER_POOL_LIMIT", CELERY_BROKER_POOL_LIMIT_DEFAULT)
+        os.environ.get("CELERY_BROKER_POOL_LIMIT", _CELERY_BROKER_POOL_LIMIT_DEFAULT)
     )
 except ValueError:
-    CELERY_BROKER_POOL_LIMIT = CELERY_BROKER_POOL_LIMIT_DEFAULT
+    CELERY_BROKER_POOL_LIMIT = _CELERY_BROKER_POOL_LIMIT_DEFAULT
 
-CELERY_WORKER_LIGHT_CONCURRENCY_DEFAULT = 24
+_CELERY_WORKER_LIGHT_CONCURRENCY_DEFAULT = 24
 try:
     CELERY_WORKER_LIGHT_CONCURRENCY = int(
         os.environ.get(
-            "CELERY_WORKER_LIGHT_CONCURRENCY", CELERY_WORKER_LIGHT_CONCURRENCY_DEFAULT
+            "CELERY_WORKER_LIGHT_CONCURRENCY",
+            _CELERY_WORKER_LIGHT_CONCURRENCY_DEFAULT,
         )
     )
 except ValueError:
-    CELERY_WORKER_LIGHT_CONCURRENCY = CELERY_WORKER_LIGHT_CONCURRENCY_DEFAULT
+    CELERY_WORKER_LIGHT_CONCURRENCY = _CELERY_WORKER_LIGHT_CONCURRENCY_DEFAULT
 
-CELERY_WORKER_LIGHT_PREFETCH_MULTIPLIER_DEFAULT = 8
+_CELERY_WORKER_LIGHT_PREFETCH_MULTIPLIER_DEFAULT = 8
 try:
     CELERY_WORKER_LIGHT_PREFETCH_MULTIPLIER = int(
         os.environ.get(
             "CELERY_WORKER_LIGHT_PREFETCH_MULTIPLIER",
-            CELERY_WORKER_LIGHT_PREFETCH_MULTIPLIER_DEFAULT,
+            _CELERY_WORKER_LIGHT_PREFETCH_MULTIPLIER_DEFAULT,
         )
     )
 except ValueError:
     CELERY_WORKER_LIGHT_PREFETCH_MULTIPLIER = (
-        CELERY_WORKER_LIGHT_PREFETCH_MULTIPLIER_DEFAULT
+        _CELERY_WORKER_LIGHT_PREFETCH_MULTIPLIER_DEFAULT
     )
 
-CELERY_WORKER_DOCPROCESSING_CONCURRENCY_DEFAULT = 6
+_CELERY_WORKER_DOCPROCESSING_CONCURRENCY_DEFAULT = 6
 try:
     env_value = os.environ.get("CELERY_WORKER_DOCPROCESSING_CONCURRENCY")
     if not env_value:
         env_value = os.environ.get("NUM_INDEXING_WORKERS")
 
     if not env_value:
-        env_value = str(CELERY_WORKER_DOCPROCESSING_CONCURRENCY_DEFAULT)
+        env_value = str(_CELERY_WORKER_DOCPROCESSING_CONCURRENCY_DEFAULT)
     CELERY_WORKER_DOCPROCESSING_CONCURRENCY = int(env_value)
 except ValueError:
     CELERY_WORKER_DOCPROCESSING_CONCURRENCY = (
-        CELERY_WORKER_DOCPROCESSING_CONCURRENCY_DEFAULT
+        _CELERY_WORKER_DOCPROCESSING_CONCURRENCY_DEFAULT
     )
 
-CELERY_WORKER_DOCFETCHING_CONCURRENCY_DEFAULT = 1
+_CELERY_WORKER_DOCFETCHING_CONCURRENCY_DEFAULT = 1
 try:
     env_value = os.environ.get("CELERY_WORKER_DOCFETCHING_CONCURRENCY")
     if not env_value:
         env_value = os.environ.get("NUM_DOCFETCHING_WORKERS")
 
     if not env_value:
-        env_value = str(CELERY_WORKER_DOCFETCHING_CONCURRENCY_DEFAULT)
+        env_value = str(_CELERY_WORKER_DOCFETCHING_CONCURRENCY_DEFAULT)
     CELERY_WORKER_DOCFETCHING_CONCURRENCY = int(env_value)
 except ValueError:
     CELERY_WORKER_DOCFETCHING_CONCURRENCY = (
-        CELERY_WORKER_DOCFETCHING_CONCURRENCY_DEFAULT
+        _CELERY_WORKER_DOCFETCHING_CONCURRENCY_DEFAULT
     )
 
 CELERY_WORKER_PRIMARY_CONCURRENCY = int(
@@ -469,11 +478,6 @@ CONFLUENCE_CONNECTOR_LABELS_TO_SKIP = [
     if ignored_tag
 ]
 
-# Avoid to get archived pages
-CONFLUENCE_CONNECTOR_INDEX_ARCHIVED_PAGES = (
-    os.environ.get("CONFLUENCE_CONNECTOR_INDEX_ARCHIVED_PAGES", "").lower() == "true"
-)
-
 # Attachments exceeding this size will not be retrieved (in bytes)
 CONFLUENCE_CONNECTOR_ATTACHMENT_SIZE_THRESHOLD = int(
     os.environ.get("CONFLUENCE_CONNECTOR_ATTACHMENT_SIZE_THRESHOLD", 10 * 1024 * 1024)
@@ -598,14 +602,6 @@ MAX_SLACK_THREAD_CONTEXT_MESSAGES = int(
     os.environ.get("MAX_SLACK_THREAD_CONTEXT_MESSAGES", "5")
 )
 
-DASK_JOB_CLIENT_ENABLED = (
-    os.environ.get("DASK_JOB_CLIENT_ENABLED", "").lower() == "true"
-)
-EXPERIMENTAL_CHECKPOINTING_ENABLED = (
-    os.environ.get("EXPERIMENTAL_CHECKPOINTING_ENABLED", "").lower() == "true"
-)
-
-
 # TestRail specific configs
 TESTRAIL_BASE_URL = os.environ.get("TESTRAIL_BASE_URL", "")
 TESTRAIL_USERNAME = os.environ.get("TESTRAIL_USERNAME", "")
@@ -616,7 +612,6 @@ LEAVE_CONNECTOR_ACTIVE_ON_INITIALIZATION_FAILURE = (
     == "true"
 )
 
-PRUNING_DISABLED = -1
 DEFAULT_PRUNING_FREQ = 60 * 60 * 24  # Once a day
 
 ALLOW_SIMULTANEOUS_PRUNING = (
@@ -668,10 +663,6 @@ LARGE_CHUNK_RATIO = 4
 # Include the document level metadata in each chunk. If the metadata is too long, then it is thrown out
 # We don't want the metadata to overwhelm the actual contents of the chunk
 SKIP_METADATA_IN_CHUNK = os.environ.get("SKIP_METADATA_IN_CHUNK", "").lower() == "true"
-# Timeout to wait for job's last update before killing it, in hours
-CLEANUP_INDEXING_JOBS_TIMEOUT = int(
-    os.environ.get("CLEANUP_INDEXING_JOBS_TIMEOUT") or 3
-)
 
 # The indexer will warn in the logs whenver a document exceeds this threshold (in bytes)
 INDEXING_SIZE_WARNING_THRESHOLD = int(
@@ -688,14 +679,6 @@ INDEXING_EMBEDDING_MODEL_NUM_THREADS = int(
     os.environ.get("INDEXING_EMBEDDING_MODEL_NUM_THREADS") or 8
 )
 
-# During an indexing attempt, specifies the number of batches which are allowed to
-# exception without aborting the attempt.
-INDEXING_EXCEPTION_LIMIT = int(os.environ.get("INDEXING_EXCEPTION_LIMIT") or 0)
-
-# Maximum number of user file connector credential pairs to index in a single batch
-# Setting this number too high may overload the indexing process
-USER_FILE_INDEXING_LIMIT = int(os.environ.get("USER_FILE_INDEXING_LIMIT") or 100)
-
 # Maximum file size in a document to be indexed
 MAX_DOCUMENT_CHARS = int(os.environ.get("MAX_DOCUMENT_CHARS") or 5_000_000)
 MAX_FILE_SIZE_BYTES = int(
@@ -742,22 +725,10 @@ CODE_INTERPRETER_MAX_OUTPUT_LENGTH = int(
 # Miscellaneous
 #####
 JOB_TIMEOUT = 60 * 60 * 6  # 6 hours default
-# used to allow the background indexing jobs to use a different embedding
-# model server than the API server
-CURRENT_PROCESS_IS_AN_INDEXING_JOB = (
-    os.environ.get("CURRENT_PROCESS_IS_AN_INDEXING_JOB", "").lower() == "true"
-)
-# Sets LiteLLM to verbose logging
-LOG_ALL_MODEL_INTERACTIONS = (
-    os.environ.get("LOG_ALL_MODEL_INTERACTIONS", "").lower() == "true"
-)
 # Logs Onyx only model interactions like prompts, responses, messages etc.
 LOG_ONYX_MODEL_INTERACTIONS = (
     os.environ.get("LOG_ONYX_MODEL_INTERACTIONS", "").lower() == "true"
 )
-LOG_INDIVIDUAL_MODEL_TOKENS = (
-    os.environ.get("LOG_INDIVIDUAL_MODEL_TOKENS", "").lower() == "true"
-)
 # If set to `true` will enable additional logs about Vespa query performance
 # (time spent on finding the right docs + time spent fetching summaries from disk)
 LOG_VESPA_TIMING_INFORMATION = (
@@ -779,7 +750,27 @@ BRAINTRUST_PROJECT = os.environ.get("BRAINTRUST_PROJECT", "Onyx")
 # Braintrust API key - if provided, Braintrust tracing will be enabled
 BRAINTRUST_API_KEY = os.environ.get("BRAINTRUST_API_KEY") or ""
 # Maximum concurrency for Braintrust evaluations
-BRAINTRUST_MAX_CONCURRENCY = int(os.environ.get("BRAINTRUST_MAX_CONCURRENCY") or 5)
+# None means unlimited concurrency, otherwise specify a number
+_braintrust_concurrency = os.environ.get("BRAINTRUST_MAX_CONCURRENCY")
+BRAINTRUST_MAX_CONCURRENCY = (
+    int(_braintrust_concurrency) if _braintrust_concurrency else None
+)
+
+#####
+# Scheduled Evals Configuration
+#####
+# Comma-separated list of Braintrust dataset names to run on schedule
+SCHEDULED_EVAL_DATASET_NAMES = [
+    name.strip()
+    for name in os.environ.get("SCHEDULED_EVAL_DATASET_NAMES", "").split(",")
+    if name.strip()
+]
+# Email address to use for search permissions during scheduled evals
+SCHEDULED_EVAL_PERMISSIONS_EMAIL = os.environ.get(
+    "SCHEDULED_EVAL_PERMISSIONS_EMAIL", "roshan@onyx.app"
+)
+# Braintrust project name to use for scheduled evals
+SCHEDULED_EVAL_PROJECT = os.environ.get("SCHEDULED_EVAL_PROJECT", "st-dev")
 
 #####
 # Langfuse Configuration
@@ -788,10 +779,6 @@ BRAINTRUST_MAX_CONCURRENCY = int(os.environ.get("BRAINTRUST_MAX_CONCURRENCY") or
 LANGFUSE_SECRET_KEY = os.environ.get("LANGFUSE_SECRET_KEY") or ""
 LANGFUSE_PUBLIC_KEY = os.environ.get("LANGFUSE_PUBLIC_KEY") or ""
 
-TOKEN_BUDGET_GLOBALLY_ENABLED = (
-    os.environ.get("TOKEN_BUDGET_GLOBALLY_ENABLED", "").lower() == "true"
-)
-
 # Defined custom query/answer conditions to validate the query and the LLM answer.
 # Format: list of strings
 CUSTOM_ANSWER_VALIDITY_CONDITIONS = json.loads(
@@ -819,16 +806,16 @@ try:
 except json.JSONDecodeError:
     pass
 
-# LLM Model Update API endpoint
-LLM_MODEL_UPDATE_API_URL = os.environ.get("LLM_MODEL_UPDATE_API_URL")
+# Auto LLM Configuration - fetches model configs from GitHub for providers in Auto mode
+AUTO_LLM_CONFIG_URL = os.environ.get(
+    "AUTO_LLM_CONFIG_URL",
+    "https://raw.githubusercontent.com/onyx-dot-app/onyx/main/backend/onyx/llm/well_known_providers/recommended-models.json",
+)
 
-# Federated Search Configs
-MAX_FEDERATED_SECTIONS = int(
-    os.environ.get("MAX_FEDERATED_SECTIONS", "5")
-)  # max no. of federated sections to always keep
-MAX_FEDERATED_CHUNKS = int(
-    os.environ.get("MAX_FEDERATED_CHUNKS", "5")
-)  # max no. of chunks to retrieve per federated connector
+# How often to check for auto LLM model updates (in seconds)
+AUTO_LLM_UPDATE_INTERVAL_SECONDS = int(
+    os.environ.get("AUTO_LLM_UPDATE_INTERVAL_SECONDS", 1800)  # 30 minutes
+)
 
 #####
 # Enterprise Edition Configs
@@ -841,6 +828,11 @@ ENTERPRISE_EDITION_ENABLED = (
     os.environ.get("ENABLE_PAID_ENTERPRISE_EDITION_FEATURES", "").lower() == "true"
 )
 
+#####
+# Image Generation Configuration (DEPRECATED)
+# These environment variables will be deprecated soon.
+# To configure image generation, please visit the Image Generation page in the Admin Panel.
+#####
 # Azure Image Configurations
 AZURE_IMAGE_API_VERSION = os.environ.get("AZURE_IMAGE_API_VERSION") or os.environ.get(
     "AZURE_DALLE_API_VERSION"
@@ -884,8 +876,6 @@ OAUTH_CONFLUENCE_CLOUD_CLIENT_ID = os.environ.get(
 OAUTH_CONFLUENCE_CLOUD_CLIENT_SECRET = os.environ.get(
     "OAUTH_CONFLUENCE_CLOUD_CLIENT_SECRET", ""
 )
-OAUTH_JIRA_CLOUD_CLIENT_ID = os.environ.get("OAUTH_JIRA_CLOUD_CLIENT_ID", "")
-OAUTH_JIRA_CLOUD_CLIENT_SECRET = os.environ.get("OAUTH_JIRA_CLOUD_CLIENT_SECRET", "")
 OAUTH_GOOGLE_DRIVE_CLIENT_ID = os.environ.get("OAUTH_GOOGLE_DRIVE_CLIENT_ID", "")
 OAUTH_GOOGLE_DRIVE_CLIENT_SECRET = os.environ.get(
     "OAUTH_GOOGLE_DRIVE_CLIENT_SECRET", ""
@@ -927,9 +917,20 @@ DEV_MODE = os.environ.get("DEV_MODE", "").lower() == "true"
 
 INTEGRATION_TESTS_MODE = os.environ.get("INTEGRATION_TESTS_MODE", "").lower() == "true"
 
-MOCK_CONNECTOR_FILE_PATH = os.environ.get("MOCK_CONNECTOR_FILE_PATH")
+#####
+# Captcha Configuration (for cloud signup protection)
+#####
+# Enable captcha verification for new user registration
+CAPTCHA_ENABLED = os.environ.get("CAPTCHA_ENABLED", "").lower() == "true"
 
-TEST_ENV = os.environ.get("TEST_ENV", "").lower() == "true"
+# Google reCAPTCHA secret key (server-side validation)
+RECAPTCHA_SECRET_KEY = os.environ.get("RECAPTCHA_SECRET_KEY", "")
+
+# Minimum score threshold for reCAPTCHA v3 (0.0-1.0, higher = more likely human)
+# 0.5 is the recommended default
+RECAPTCHA_SCORE_THRESHOLD = float(os.environ.get("RECAPTCHA_SCORE_THRESHOLD", "0.5"))
+
+MOCK_CONNECTOR_FILE_PATH = os.environ.get("MOCK_CONNECTOR_FILE_PATH")
 
 # Set to true to mock LLM responses for testing purposes
 MOCK_LLM_RESPONSE = (
@@ -955,15 +956,6 @@ IMAGE_SUMMARIZATION_USER_PROMPT = os.environ.get(
     DEFAULT_IMAGE_SUMMARIZATION_USER_PROMPT,
 )
 
-IMAGE_ANALYSIS_SYSTEM_PROMPT = os.environ.get(
-    "IMAGE_ANALYSIS_SYSTEM_PROMPT",
-    DEFAULT_IMAGE_ANALYSIS_SYSTEM_PROMPT,
-)
-
-DISABLE_AUTO_AUTH_REFRESH = (
-    os.environ.get("DISABLE_AUTO_AUTH_REFRESH", "").lower() == "true"
-)
-
 # Knowledge Graph Read Only User Configuration
 DB_READONLY_USER: str = os.environ.get("DB_READONLY_USER", "db_readonly_user")
 DB_READONLY_PASSWORD: str = urllib.parse.quote_plus(
@@ -991,3 +983,15 @@ S3_GENERATE_LOCAL_CHECKSUM = (
 # Forcing Vespa Language
 # English: en, German:de, etc. See: https://docs.vespa.ai/en/linguistics.html
 VESPA_LANGUAGE_OVERRIDE = os.environ.get("VESPA_LANGUAGE_OVERRIDE")
+
+
+#####
+# Default LLM API Keys (for cloud deployments)
+# These are Onyx-managed API keys provided to tenants by default
+#####
+OPENAI_DEFAULT_API_KEY = os.environ.get("OPENAI_DEFAULT_API_KEY")
+ANTHROPIC_DEFAULT_API_KEY = os.environ.get("ANTHROPIC_DEFAULT_API_KEY")
+COHERE_DEFAULT_API_KEY = os.environ.get("COHERE_DEFAULT_API_KEY")
+VERTEXAI_DEFAULT_CREDENTIALS = os.environ.get("VERTEXAI_DEFAULT_CREDENTIALS")
+VERTEXAI_DEFAULT_LOCATION = os.environ.get("VERTEXAI_DEFAULT_LOCATION", "global")
+OPENROUTER_DEFAULT_API_KEY = os.environ.get("OPENROUTER_DEFAULT_API_KEY")

backend/onyx/configs/chat_configs.py

@@ -11,9 +11,6 @@ NUM_POSTPROCESSED_RESULTS = 20
 
 # May be less depending on model
 MAX_CHUNKS_FED_TO_CHAT = int(os.environ.get("MAX_CHUNKS_FED_TO_CHAT") or 25)
-# For Chat, need to keep enough space for history and other prompt pieces
-# ~3k input, half for docs, half for chat history + prompts
-CHAT_TARGET_CHUNK_PERCENTAGE = 512 * 3 / 3072
 
 # Maximum percentage of the context window to fill with selected sections
 SELECTED_SECTIONS_MAX_WINDOW_PERCENTAGE = 0.8

backend/onyx/configs/constants.py

@@ -146,9 +146,6 @@ CELERY_PERMISSIONS_SYNC_LOCK_TIMEOUT = 3600  # 1 hour (in seconds)
 
 CELERY_EXTERNAL_GROUP_SYNC_LOCK_TIMEOUT = 300  # 5 min
 
-# Doc ID migration can be long-running; use a longer TTL and renew periodically
-CELERY_USER_FILE_DOCID_MIGRATION_LOCK_TIMEOUT = 10 * 60  # 10 minutes (in seconds)
-
 CELERY_USER_FILE_PROCESSING_LOCK_TIMEOUT = 30 * 60  # 30 minutes (in seconds)
 
 CELERY_USER_FILE_PROJECT_SYNC_LOCK_TIMEOUT = 5 * 60  # 5 minutes (in seconds)
@@ -237,6 +234,7 @@ class NotificationType(str, Enum):
     REINDEX = "reindex"
     PERSONA_SHARED = "persona_shared"
     TRIAL_ENDS_TWO_DAYS = "two_day_trial_ending"  # 2 days left in trial
+    RELEASE_NOTES = "release_notes"
 
 
 class BlobType(str, Enum):
@@ -365,9 +363,6 @@ class OnyxCeleryQueues:
     CONNECTOR_EXTERNAL_GROUP_SYNC = "connector_external_group_sync"
     CSV_GENERATION = "csv_generation"
 
-    # Indexing queue
-    USER_FILES_INDEXING = "user_files_indexing"
-
     # User file processing queue
     USER_FILE_PROCESSING = "user_file_processing"
     USER_FILE_PROJECT_SYNC = "user_file_project_sync"
@@ -426,7 +421,6 @@ class OnyxRedisLocks:
     USER_FILE_PROJECT_SYNC_LOCK_PREFIX = "da_lock:user_file_project_sync"
     USER_FILE_DELETE_BEAT_LOCK = "da_lock:check_user_file_delete_beat"
     USER_FILE_DELETE_LOCK_PREFIX = "da_lock:user_file_delete"
-    USER_FILE_DOCID_MIGRATION_LOCK = "da_lock:user_file_docid_migration"
 
 
 class OnyxRedisSignals:
@@ -492,7 +486,7 @@ class OnyxCeleryTask:
     CHECK_FOR_PRUNING = "check_for_pruning"
     CHECK_FOR_DOC_PERMISSIONS_SYNC = "check_for_doc_permissions_sync"
     CHECK_FOR_EXTERNAL_GROUP_SYNC = "check_for_external_group_sync"
-    CHECK_FOR_LLM_MODEL_UPDATE = "check_for_llm_model_update"
+    CHECK_FOR_AUTO_LLM_UPDATE = "check_for_auto_llm_update"
 
     # User file processing
     CHECK_FOR_USER_FILE_PROCESSING = "check_for_user_file_processing"
@@ -533,7 +527,6 @@ class OnyxCeleryTask:
     CONNECTOR_PRUNING_GENERATOR_TASK = "connector_pruning_generator_task"
     DOCUMENT_BY_CC_PAIR_CLEANUP_TASK = "document_by_cc_pair_cleanup_task"
     VESPA_METADATA_SYNC_TASK = "vespa_metadata_sync_task"
-    USER_FILE_DOCID_MIGRATION = "user_file_docid_migration"
 
     # chat retention
     CHECK_TTL_MANAGEMENT_TASK = "check_ttl_management_task"
@@ -542,6 +535,7 @@ class OnyxCeleryTask:
     GENERATE_USAGE_REPORT_TASK = "generate_usage_report_task"
 
     EVAL_RUN_TASK = "eval_run_task"
+    SCHEDULED_EVAL_TASK = "scheduled_eval_task"
 
     EXPORT_QUERY_HISTORY_TASK = "export_query_history_task"
     EXPORT_QUERY_HISTORY_CLEANUP_TASK = "export_query_history_cleanup_task"
@@ -562,9 +556,9 @@ REDIS_SOCKET_KEEPALIVE_OPTIONS[socket.TCP_KEEPINTVL] = 15
 REDIS_SOCKET_KEEPALIVE_OPTIONS[socket.TCP_KEEPCNT] = 3
 
 if platform.system() == "Darwin":
-    REDIS_SOCKET_KEEPALIVE_OPTIONS[socket.TCP_KEEPALIVE] = 60  # type: ignore
+    REDIS_SOCKET_KEEPALIVE_OPTIONS[socket.TCP_KEEPALIVE] = 60  # type: ignore[attr-defined,unused-ignore]
 else:
-    REDIS_SOCKET_KEEPALIVE_OPTIONS[socket.TCP_KEEPIDLE] = 60
+    REDIS_SOCKET_KEEPALIVE_OPTIONS[socket.TCP_KEEPIDLE] = 60  # type: ignore[attr-defined,unused-ignore]
 
 
 class OnyxCallTypes(str, Enum):

backend/onyx/configs/model_configs.py

@@ -78,6 +78,10 @@ GEN_AI_MODEL_FALLBACK_MAX_TOKENS = int(
 GEN_AI_SINGLE_USER_MESSAGE_EXPECTED_MAX_TOKENS = 512
 GEN_AI_TEMPERATURE = float(os.environ.get("GEN_AI_TEMPERATURE") or 0)
 
+# Reasoning models use effort to control the amount of reasoning
+# before tool calling or answer generation.
+DEFAULT_REASONING_EFFORT = os.environ.get("DEFAULT_REASONING_EFFORT") or "low"
+
 # should be used if you are using a custom LLM inference provider that doesn't support
 # streaming format AND you are still using the langchain/litellm LLM class
 DISABLE_LITELLM_STREAMING = (
@@ -125,9 +129,16 @@ if _LITELLM_EXTRA_BODY_RAW:
     except Exception:
         pass
 
-# Whether and how to lower scores for short chunks w/o relevant context
-# Evaluated via custom ML model
-
-USE_INFORMATION_CONTENT_CLASSIFICATION = (
-    os.environ.get("USE_INFORMATION_CONTENT_CLASSIFICATION", "false").lower() == "true"
+#####
+# Prompt Caching Configs
+#####
+# Enable prompt caching framework
+ENABLE_PROMPT_CACHING = (
+    os.environ.get("ENABLE_PROMPT_CACHING", "true").lower() != "false"
+)
+
+# Cache TTL multiplier - store caches slightly longer than provider TTL
+# This allows for some clock skew and ensures we don't lose cache metadata prematurely
+PROMPT_CACHE_REDIS_TTL_MULTIPLIER = float(
+    os.environ.get("PROMPT_CACHE_REDIS_TTL_MULTIPLIER") or 1.2
 )

backend/onyx/connectors/confluence/onyx_confluence.py

@@ -961,14 +961,20 @@ def get_user_email_from_username__server(
         try:
             response = confluence_client.get_mobile_parameters(user_name)
             email = response.get("email")
-        except Exception:
-            logger.warning(f"failed to get confluence email for {user_name}")
+        except HTTPError as e:
+            status_code = e.response.status_code if e.response is not None else "N/A"
+            logger.warning(
+                f"Failed to get confluence email for {user_name}: "
+                f"HTTP {status_code} - {e}"
+            )
             # For now, we'll just return None and log a warning. This means
             # we will keep retrying to get the email every group sync.
             email = None
-            # We may want to just return a string that indicates failure so we dont
-            # keep retrying
-            # email = f"FAILED TO GET CONFLUENCE EMAIL FOR {user_name}"
+        except Exception as e:
+            logger.warning(
+                f"Failed to get confluence email for {user_name}: {type(e).__name__} - {e}"
+            )
+            email = None
         _USER_EMAIL_CACHE[user_name] = email
     return _USER_EMAIL_CACHE[user_name]
 

backend/onyx/connectors/drupal_wiki/connector.py

@@ -58,51 +58,59 @@ class DrupalWikiConnector(
     CheckpointedConnector[DrupalWikiCheckpoint],
     SlimConnector,
 ):
+    # Deprecated parameters that may exist in old connector configurations
+    _DEPRECATED_PARAMS = {"drupal_wiki_scope", "include_all_spaces"}
+
     def __init__(
         self,
         base_url: str,
         spaces: list[str] | None = None,
         pages: list[str] | None = None,
-        include_all_spaces: bool = False,
         batch_size: int = INDEX_BATCH_SIZE,
         continue_on_failure: bool = CONTINUE_ON_CONNECTOR_FAILURE,
-        drupal_wiki_scope: str | None = None,
         include_attachments: bool = False,
         allow_images: bool = False,
+        **kwargs: Any,
     ) -> None:
         """
         Initialize the Drupal Wiki connector.
 
         Args:
             base_url: The base URL of the Drupal Wiki instance (e.g., https://help.drupal-wiki.com)
-            spaces: List of space IDs to index. If None and include_all_spaces is False, no spaces will be indexed.
-            pages: List of page IDs to index. If provided, only these specific pages will be indexed.
-            include_all_spaces: If True, all spaces will be indexed regardless of the spaces parameter.
+            spaces: List of space IDs to index. If None and pages is also None, all spaces will be indexed.
+            pages: List of page IDs to index. If provided, these specific pages will be indexed.
             batch_size: Number of documents to process in a batch.
             continue_on_failure: If True, continue indexing even if some documents fail.
-            drupal_wiki_scope: The selected tab value from the frontend. If "all_spaces", all spaces will be indexed.
             include_attachments: If True, enable processing of page attachments including images and documents.
             allow_images: If True, enable processing of image attachments.
         """
+
+        #########################################################
+        # TODO: Remove this after 02/01/2026 and remove **kwargs from the function signature
+        # Check for deprecated parameters from old connector configurations
+        # If attempting to update without deleting the connector:
+        # Remove the deprecated parameters from the custom_connector_config in the relevant connector table rows
+        deprecated_found = set(kwargs.keys()) & self._DEPRECATED_PARAMS
+        if deprecated_found:
+            raise ConnectorValidationError(
+                f"Outdated Drupal Wiki connector configuration detected "
+                f"(found deprecated parameters: {', '.join(deprecated_found)}). "
+                f"Please delete and recreate this connector, or contact Onyx support "
+                f"for assistance with updating the configuration without deleting the connector."
+            )
+        # Reject any other unexpected parameters
+        if kwargs:
+            raise ConnectorValidationError(
+                f"Unexpected parameters for Drupal Wiki connector: {', '.join(kwargs.keys())}"
+            )
+        #########################################################
+
         self.base_url = base_url.rstrip("/")
         self.spaces = spaces or []
         self.pages = pages or []
 
-        # Determine whether to include all spaces based on the selected tab
-        # If drupal_wiki_scope is "all_spaces", we should index all spaces
-        # If it's "specific_spaces", we should only index the specified spaces
-        # If it's None, we use the include_all_spaces parameter
-
-        if drupal_wiki_scope is not None:
-            logger.debug(f"drupal_wiki_scope is set to {drupal_wiki_scope}")
-
-            self.include_all_spaces = drupal_wiki_scope == "all_spaces"
-            # If scope is specific_spaces, include_all_spaces correctly defaults to False
-        else:
-            logger.debug(
-                f"drupal_wiki_scope is not set, using include_all_spaces={include_all_spaces}"
-            )
-            self.include_all_spaces = include_all_spaces
+        # If no specific spaces or pages are provided, index all spaces
+        self.include_all_spaces = not self.spaces and not self.pages
 
         self.batch_size = batch_size
         self.continue_on_failure = continue_on_failure

backend/onyx/connectors/google_drive/connector.py

@@ -12,7 +12,9 @@ from functools import partial
 from typing import Any
 from typing import cast
 from typing import Protocol
+from urllib.parse import parse_qs
 from urllib.parse import urlparse
+from urllib.parse import urlunparse
 
 from google.auth.exceptions import RefreshError
 from google.oauth2.credentials import Credentials as OAuthCredentials
@@ -64,6 +66,7 @@ from onyx.connectors.google_utils.shared_constants import USER_FIELDS
 from onyx.connectors.interfaces import CheckpointedConnectorWithPermSync
 from onyx.connectors.interfaces import CheckpointOutput
 from onyx.connectors.interfaces import GenerateSlimDocumentOutput
+from onyx.connectors.interfaces import NormalizationResult
 from onyx.connectors.interfaces import SecondsSinceUnixEpoch
 from onyx.connectors.interfaces import SlimConnectorWithPermSync
 from onyx.connectors.models import ConnectorFailure
@@ -281,6 +284,54 @@ class GoogleDriveConnector(
             )
         return self._creds
 
+    @classmethod
+    @override
+    def normalize_url(cls, url: str) -> NormalizationResult:
+        """Normalize a Google Drive URL to match the canonical Document.id format.
+
+        Reuses the connector's existing document ID creation logic from
+        onyx_document_id_from_drive_file.
+        """
+        parsed = urlparse(url)
+        netloc = parsed.netloc.lower()
+
+        if not (
+            netloc.startswith("docs.google.com")
+            or netloc.startswith("drive.google.com")
+        ):
+            return NormalizationResult(normalized_url=None, use_default=False)
+
+        # Handle ?id= query parameter case
+        query_params = parse_qs(parsed.query)
+        doc_id = query_params.get("id", [None])[0]
+        if doc_id:
+            scheme = parsed.scheme or "https"
+            netloc = "drive.google.com"
+            path = f"/file/d/{doc_id}"
+            params = ""
+            query = ""
+            fragment = ""
+            normalized = urlunparse(
+                (scheme, netloc, path, params, query, fragment)
+            ).rstrip("/")
+            return NormalizationResult(normalized_url=normalized, use_default=False)
+
+        # Extract file ID and use connector's function
+        path_parts = parsed.path.split("/")
+        file_id = None
+        for i, part in enumerate(path_parts):
+            if part == "d" and i + 1 < len(path_parts):
+                file_id = path_parts[i + 1]
+                break
+
+        if not file_id:
+            return NormalizationResult(normalized_url=None, use_default=False)
+
+        # Create minimal file object for connector function
+        file_obj = {"webViewLink": url, "id": file_id}
+        normalized = onyx_document_id_from_drive_file(file_obj).rstrip("/")
+        return NormalizationResult(normalized_url=normalized, use_default=False)
+
     # TODO: ensure returned new_creds_dict is actually persisted when this is called?
     def load_credentials(self, credentials: dict[str, Any]) -> dict[str, str] | None:
         try:

backend/onyx/connectors/interfaces.py

@@ -25,6 +25,18 @@ GenerateSlimDocumentOutput = Iterator[list[SlimDocument]]
 CT = TypeVar("CT", bound=ConnectorCheckpoint)
 
 
+class NormalizationResult(BaseModel):
+    """Result of URL normalization attempt.
+
+    Attributes:
+        normalized_url: The normalized URL string, or None if normalization failed
+        use_default: If True, fall back to default normalizer. If False, return None.
+    """
+
+    normalized_url: str | None
+    use_default: bool = False
+
+
 class BaseConnector(abc.ABC, Generic[CT]):
     REDIS_KEY_PREFIX = "da_connector_data:"
 
@@ -74,6 +86,15 @@ class BaseConnector(abc.ABC, Generic[CT]):
         """Implement if the underlying connector wants to skip/allow image downloading
         based on the application level image analysis setting."""
 
+    @classmethod
+    def normalize_url(cls, url: str) -> "NormalizationResult":
+        """Normalize a URL to match the canonical Document.id format used during ingestion.
+
+        Connectors that use URLs as document IDs should override this method.
+        Returns NormalizationResult with use_default=True if not implemented.
+        """
+        return NormalizationResult(normalized_url=None, use_default=True)
+
     def build_dummy_checkpoint(self) -> CT:
         # TODO: find a way to make this work without type: ignore
         return ConnectorCheckpoint(has_more=True)  # type: ignore

backend/onyx/connectors/linear/connector.py

@@ -1,10 +1,13 @@
 import os
+import re
 from datetime import datetime
 from datetime import timezone
 from typing import Any
 from typing import cast
+from urllib.parse import urlparse
 
 import requests
+from typing_extensions import override
 
 from onyx.configs.app_configs import INDEX_BATCH_SIZE
 from onyx.configs.app_configs import LINEAR_CLIENT_ID
@@ -16,6 +19,7 @@ from onyx.connectors.cross_connector_utils.miscellaneous_utils import (
 from onyx.connectors.cross_connector_utils.miscellaneous_utils import time_str_to_utc
 from onyx.connectors.interfaces import GenerateDocumentsOutput
 from onyx.connectors.interfaces import LoadConnector
+from onyx.connectors.interfaces import NormalizationResult
 from onyx.connectors.interfaces import OAuthConnector
 from onyx.connectors.interfaces import PollConnector
 from onyx.connectors.interfaces import SecondsSinceUnixEpoch
@@ -312,6 +316,31 @@ class LinearConnector(LoadConnector, PollConnector, OAuthConnector):
 
         yield from self._process_issues(start_str=start_time, end_str=end_time)
 
+    @classmethod
+    @override
+    def normalize_url(cls, url: str) -> NormalizationResult:
+        """Extract Linear issue identifier from URL.
+
+        Linear URLs are like: https://linear.app/team/issue/IDENTIFIER/...
+        Returns the identifier (e.g., "DAN-2327") which can be used to match Document.link.
+        """
+        parsed = urlparse(url)
+        netloc = parsed.netloc.lower()
+
+        if "linear.app" not in netloc:
+            return NormalizationResult(normalized_url=None, use_default=False)
+
+        # Extract identifier from path: /team/issue/IDENTIFIER/...
+        # Pattern: /{team}/issue/{identifier}/...
+        path_parts = [p for p in parsed.path.split("/") if p]
+        if len(path_parts) >= 3 and path_parts[1] == "issue":
+            identifier = path_parts[2]
+            # Validate identifier format (e.g., "DAN-2327")
+            if re.match(r"^[A-Z]+-\d+$", identifier):
+                return NormalizationResult(normalized_url=identifier, use_default=False)
+
+        return NormalizationResult(normalized_url=None, use_default=False)
+
 
 if __name__ == "__main__":
     connector = LinearConnector()

backend/onyx/connectors/notion/connector.py

@@ -1,13 +1,17 @@
+import re
 from collections.abc import Generator
 from datetime import datetime
 from datetime import timezone
 from typing import Any
 from typing import cast
 from typing import Optional
+from urllib.parse import parse_qs
+from urllib.parse import urlparse
 
 import requests
 from pydantic import BaseModel
 from retry import retry
+from typing_extensions import override
 
 from onyx.configs.app_configs import INDEX_BATCH_SIZE
 from onyx.configs.app_configs import NOTION_CONNECTOR_DISABLE_RECURSIVE_PAGE_LOOKUP
@@ -21,6 +25,7 @@ from onyx.connectors.exceptions import InsufficientPermissionsError
 from onyx.connectors.exceptions import UnexpectedValidationError
 from onyx.connectors.interfaces import GenerateDocumentsOutput
 from onyx.connectors.interfaces import LoadConnector
+from onyx.connectors.interfaces import NormalizationResult
 from onyx.connectors.interfaces import PollConnector
 from onyx.connectors.interfaces import SecondsSinceUnixEpoch
 from onyx.connectors.models import ConnectorMissingCredentialError
@@ -101,6 +106,49 @@ class NotionConnector(LoadConnector, PollConnector):
         # very large, this may not be practical.
         self.recursive_index_enabled = recursive_index_enabled or self.root_page_id
 
+    @classmethod
+    @override
+    def normalize_url(cls, url: str) -> NormalizationResult:
+        """Normalize a Notion URL to extract the page ID (UUID format)."""
+        parsed = urlparse(url)
+        netloc = parsed.netloc.lower()
+
+        if not ("notion.so" in netloc or "notion.site" in netloc):
+            return NormalizationResult(normalized_url=None, use_default=False)
+
+        # Extract page ID from path (format: "Title-PageID")
+        path_last = parsed.path.split("/")[-1]
+        candidate = path_last.split("-")[-1] if "-" in path_last else path_last
+
+        # Clean and format as UUID
+        candidate = re.sub(r"[^0-9a-fA-F-]", "", candidate)
+        cleaned = candidate.replace("-", "")
+
+        if len(cleaned) == 32 and re.fullmatch(r"[0-9a-fA-F]{32}", cleaned):
+            normalized_uuid = (
+                f"{cleaned[0:8]}-{cleaned[8:12]}-{cleaned[12:16]}-"
+                f"{cleaned[16:20]}-{cleaned[20:]}"
+            ).lower()
+            return NormalizationResult(
+                normalized_url=normalized_uuid, use_default=False
+            )
+
+        # Try query params
+        params = parse_qs(parsed.query)
+        for key in ("p", "page_id"):
+            if key in params and params[key]:
+                candidate = params[key][0].replace("-", "")
+                if len(candidate) == 32 and re.fullmatch(r"[0-9a-fA-F]{32}", candidate):
+                    normalized_uuid = (
+                        f"{candidate[0:8]}-{candidate[8:12]}-{candidate[12:16]}-"
+                        f"{candidate[16:20]}-{candidate[20:]}"
+                    ).lower()
+                    return NormalizationResult(
+                        normalized_url=normalized_uuid, use_default=False
+                    )
+
+        return NormalizationResult(normalized_url=None, use_default=False)
+
     @retry(tries=3, delay=1, backoff=2)
     def _fetch_child_blocks(
         self, block_id: str, cursor: str | None = None

backend/onyx/connectors/slack/connector.py

@@ -15,6 +15,7 @@ from http.client import RemoteDisconnected
 from typing import Any
 from typing import cast
 from urllib.error import URLError
+from urllib.parse import urlparse
 
 from pydantic import BaseModel
 from redis import Redis
@@ -41,6 +42,7 @@ from onyx.connectors.interfaces import CheckpointOutput
 from onyx.connectors.interfaces import CredentialsConnector
 from onyx.connectors.interfaces import CredentialsProviderInterface
 from onyx.connectors.interfaces import GenerateSlimDocumentOutput
+from onyx.connectors.interfaces import NormalizationResult
 from onyx.connectors.interfaces import SecondsSinceUnixEpoch
 from onyx.connectors.interfaces import SlimConnectorWithPermSync
 from onyx.connectors.models import BasicExpertInfo
@@ -626,6 +628,43 @@ class SlackConnector(
         # self.delay_lock: str | None = None  # the redis key for the shared lock
         # self.delay_key: str | None = None  # the redis key for the shared delay
 
+    @classmethod
+    @override
+    def normalize_url(cls, url: str) -> NormalizationResult:
+        """Normalize a Slack URL to extract channel_id__thread_ts format."""
+        parsed = urlparse(url)
+        if "slack.com" not in parsed.netloc.lower():
+            return NormalizationResult(normalized_url=None, use_default=False)
+
+        # Slack document IDs are format: channel_id__thread_ts
+        # Extract from URL pattern: .../archives/{channel_id}/p{timestamp}
+        path_parts = parsed.path.split("/")
+        if "archives" not in path_parts:
+            return NormalizationResult(normalized_url=None, use_default=False)
+
+        archives_idx = path_parts.index("archives")
+        if archives_idx + 1 >= len(path_parts):
+            return NormalizationResult(normalized_url=None, use_default=False)
+
+        channel_id = path_parts[archives_idx + 1]
+        if archives_idx + 2 >= len(path_parts):
+            return NormalizationResult(normalized_url=None, use_default=False)
+
+        thread_part = path_parts[archives_idx + 2]
+        if not thread_part.startswith("p"):
+            return NormalizationResult(normalized_url=None, use_default=False)
+
+        # Convert p1234567890123456 to 1234567890.123456 format
+        timestamp_str = thread_part[1:]  # Remove 'p' prefix
+        if len(timestamp_str) == 16:
+            # Insert dot at position 10 to match canonical format
+            thread_ts = f"{timestamp_str[:10]}.{timestamp_str[10:]}"
+        else:
+            thread_ts = timestamp_str
+
+        normalized = f"{channel_id}__{thread_ts}"
+        return NormalizationResult(normalized_url=normalized, use_default=False)
+
     @staticmethod
     def make_credential_prefix(key: str) -> str:
         return f"connector:slack:credential_{key}"

backend/onyx/connectors/web/connector.py

@@ -1,4 +1,3 @@
-import io
 import ipaddress
 import random
 import socket
@@ -36,10 +35,11 @@ from onyx.connectors.interfaces import GenerateDocumentsOutput
 from onyx.connectors.interfaces import LoadConnector
 from onyx.connectors.models import Document
 from onyx.connectors.models import TextSection
-from onyx.file_processing.extract_file_text import read_pdf_file
 from onyx.file_processing.html_utils import web_html_cleanup
 from onyx.utils.logger import setup_logger
 from onyx.utils.sitemap import list_pages_for_site
+from onyx.utils.web_content import extract_pdf_text
+from onyx.utils.web_content import is_pdf_resource
 from shared_configs.configs import MULTI_TENANT
 
 logger = setup_logger()
@@ -116,16 +116,6 @@ DEFAULT_HEADERS = {
     "Sec-CH-UA-Platform": '"macOS"',
 }
 
-# Common PDF MIME types
-PDF_MIME_TYPES = [
-    "application/pdf",
-    "application/x-pdf",
-    "application/acrobat",
-    "application/vnd.pdf",
-    "text/pdf",
-    "text/x-pdf",
-]
-
 
 class WEB_CONNECTOR_VALID_SETTINGS(str, Enum):
     # Given a base site, index everything under that path
@@ -268,12 +258,6 @@ def get_internal_links(
     return internal_links
 
 
-def is_pdf_content(response: requests.Response) -> bool:
-    """Check if the response contains PDF content based on content-type header"""
-    content_type = response.headers.get("content-type", "").lower()
-    return any(pdf_type in content_type for pdf_type in PDF_MIME_TYPES)
-
-
 def start_playwright() -> Tuple[Playwright, BrowserContext]:
     playwright = sync_playwright().start()
 
@@ -529,14 +513,13 @@ class WebConnector(LoadConnector):
         head_response = requests.head(
             initial_url, headers=DEFAULT_HEADERS, allow_redirects=True
         )
-        is_pdf = is_pdf_content(head_response)
+        content_type = head_response.headers.get("content-type")
+        is_pdf = is_pdf_resource(initial_url, content_type)
 
-        if is_pdf or initial_url.lower().endswith(".pdf"):
+        if is_pdf:
             # PDF files are not checked for links
             response = requests.get(initial_url, headers=DEFAULT_HEADERS)
-            page_text, metadata, images = read_pdf_file(
-                file=io.BytesIO(response.content)
-            )
+            page_text, metadata = extract_pdf_text(response.content)
             last_modified = response.headers.get("Last-Modified")
 
             result.doc = Document(

backend/onyx/context/search/federated/slack_search.py

@@ -1016,9 +1016,15 @@ def slack_retrieval(
         for query_string in query_strings
     ]
 
-    # If include_dm is True, add additional searches without channel filters
-    # This allows searching DMs/group DMs while still searching the specified channels
-    if entities and entities.get("include_dm"):
+    # If include_dm is True AND we're not already searching all channels,
+    # add additional searches without channel filters.
+    # This allows searching DMs/group DMs while still searching the specified channels.
+    # Skip this if search_all_channels is already True (would be duplicate queries).
+    if (
+        entities
+        and entities.get("include_dm")
+        and not entities.get("search_all_channels")
+    ):
         # Create a minimal entities dict that won't add channel filters
         # This ensures we search ALL conversations (DMs, group DMs, private channels)
         # BUT we still want to exclude channels specified in exclude_channels

backend/onyx/context/search/federated/slack_search_utils.py

@@ -398,8 +398,8 @@ def extract_channel_references_from_query(query_text: str) -> set[str]:
     channel_patterns = [
         r"\bin\s+(?:the\s+)?([a-z0-9_-]+)\s+(?:slack\s+)?channels?\b",  # "in the office channel"
         r"\bfrom\s+(?:the\s+)?([a-z0-9_-]+)\s+(?:slack\s+)?channels?\b",  # "from the office channel"
-        r"\bin\s+#([a-z0-9_-]+)\b",  # "in #office"
-        r"\bfrom\s+#([a-z0-9_-]+)\b",  # "from #office"
+        r"\bin[:\s]*#([a-z0-9_-]+)\b",  # "in #office" or "in:#office"
+        r"\bfrom[:\s]*#([a-z0-9_-]+)\b",  # "from #office" or "from:#office"
     ]
 
     for pattern in channel_patterns:

backend/onyx/context/search/models.py

@@ -253,7 +253,7 @@ class RetrievalDetails(ChunkContext):
     # Use LLM to determine whether to do a retrieval or only rely on existing history
     # If the Persona is configured to not run search (0 chunks), this is bypassed
     # If no Prompt is configured, the only search results are shown, this is bypassed
-    run_search: OptionalSearchSetting = OptionalSearchSetting.ALWAYS
+    run_search: OptionalSearchSetting = OptionalSearchSetting.AUTO
     # Is this a real-time/streaming call or a question where Onyx can take more time?
     # Used to determine reranking flow
     real_time: bool = True

backend/onyx/context/search/pipeline.py

@@ -19,7 +19,6 @@ from onyx.db.models import Persona
 from onyx.db.models import User
 from onyx.document_index.interfaces import DocumentIndex
 from onyx.llm.interfaces import LLM
-from onyx.onyxbot.slack.models import SlackContext
 from onyx.secondary_llm_flows.source_filter import extract_source_filter
 from onyx.secondary_llm_flows.time_filter import extract_time_filter
 from onyx.utils.logger import setup_logger
@@ -249,8 +248,6 @@ def search_pipeline(
     db_session: Session,
     auto_detect_filters: bool = False,
     llm: LLM | None = None,
-    # Needed for federated Slack search
-    slack_context: SlackContext | None = None,
     # If a project ID is provided, it will be exclusively scoped to that project
     project_id: int | None = None,
 ) -> list[InferenceChunk]:
@@ -291,11 +288,9 @@ def search_pipeline(
 
     retrieved_chunks = search_chunks(
         query_request=query_request,
-        # Needed for federated Slack search
         user_id=user.id if user else None,
         document_index=document_index,
         db_session=db_session,
-        slack_context=slack_context,
     )
 
     # For some specific connectors like Salesforce, a user that has access to an object doesn't mean

backend/onyx/context/search/retrieval/search_runner.py

@@ -29,7 +29,6 @@ from onyx.document_index.vespa.shared_utils.utils import (
 from onyx.federated_connectors.federated_retrieval import (
     get_federated_retrieval_functions,
 )
-from onyx.onyxbot.slack.models import SlackContext
 from onyx.secondary_llm_flows.query_expansion import multilingual_query_expansion
 from onyx.utils.logger import setup_logger
 from onyx.utils.threadpool_concurrency import run_functions_tuples_in_parallel
@@ -331,7 +330,6 @@ def retrieve_chunks(
     retrieval_metrics_callback: (
         Callable[[RetrievalMetricsContainer], None] | None
     ) = None,
-    slack_context: SlackContext | None = None,
 ) -> list[InferenceChunk]:
     """Returns a list of the best chunks from an initial keyword/semantic/ hybrid search."""
 
@@ -348,8 +346,7 @@ def retrieve_chunks(
         user_id,
         list(query.filters.source_type) if query.filters.source_type else None,
         query.filters.document_set,
-        slack_context,
-        query.filters.user_file_ids,
+        user_file_ids=query.filters.user_file_ids,
     )
     federated_sources = set(
         federated_retrieval_info.source.to_non_federated_source()
@@ -460,7 +457,6 @@ def search_chunks(
     user_id: UUID | None,
     document_index: DocumentIndex,
     db_session: Session,
-    slack_context: SlackContext | None = None,
 ) -> list[InferenceChunk]:
     run_queries: list[tuple[Callable, tuple]] = []
 
@@ -476,7 +472,6 @@ def search_chunks(
         user_id=user_id,
         source_types=list(source_filters) if source_filters else None,
         document_set_names=query_request.filters.document_set,
-        slack_context=slack_context,
         user_file_ids=query_request.filters.user_file_ids,
     )
 

backend/onyx/db/README.md

@@ -27,9 +27,10 @@ Tool calls are stored in the ToolCall table and can represent all of the followi
   the agent call as a parent are the tool calls that happen as part of the agent.
 
 The different branches are generated by sending a new search query to an existing parent.
-
+```
                  [Empty Root Message]  (This allows the first message to be branched/edited as well)
               /           |           \
 [First Message] [First Message Edit 1] [First Message Edit 2]
        |                  |
 [Second Message]  [Second Message of Edit 1 Branch]
+```

backend/onyx/db/chat.py

@@ -7,6 +7,7 @@ from uuid import UUID
 from fastapi import HTTPException
 from sqlalchemy import delete
 from sqlalchemy import desc
+from sqlalchemy import exists
 from sqlalchemy import func
 from sqlalchemy import nullsfirst
 from sqlalchemy import or_
@@ -152,6 +153,7 @@ def get_chat_sessions_by_user(
     limit: int = 50,
     project_id: int | None = None,
     only_non_project_chats: bool = False,
+    include_failed_chats: bool = False,
 ) -> list[ChatSession]:
     stmt = select(ChatSession).where(ChatSession.user_id == user_id)
 
@@ -171,6 +173,16 @@ def get_chat_sessions_by_user(
     elif only_non_project_chats:
         stmt = stmt.where(ChatSession.project_id.is_(None))
 
+    if not include_failed_chats:
+        non_system_message_exists_subq = (
+            exists()
+            .where(ChatMessage.chat_session_id == ChatSession.id)
+            .where(ChatMessage.message_type != MessageType.SYSTEM)
+            .correlate(ChatSession)
+        )
+
+        stmt = stmt.where(non_system_message_exists_subq)
+
     result = db_session.execute(stmt)
     chat_sessions = result.scalars().all()