fix(review): address Greptile comments

- Add owner to bare TODO comment
- Restore placement field assertions weakened by Emitter refactor

.github/workflows/helm-chart-releases.yml

@@ -47,7 +47,8 @@ jobs:
           done
 
       - name: Publish Helm charts to gh-pages
-        uses: stefanprodan/helm-gh-pages@0ad2bb377311d61ac04ad9eb6f252fb68e207260 # ratchet:stefanprodan/helm-gh-pages@v1.7.0
+        # NOTE: HEAD of https://github.com/stefanprodan/helm-gh-pages/pull/43
+        uses: stefanprodan/helm-gh-pages@ad32ad3b8720abfeaac83532fd1e9bdfca5bbe27 # zizmor: ignore[impostor-commit]
         with:
           token: ${{ secrets.GITHUB_TOKEN }}
           charts_dir: deployment/helm/charts

.github/workflows/nightly-llm-provider-chat.yml

@@ -35,6 +35,7 @@ jobs:
     needs: [provider-chat-test]
     if: failure() && github.event_name == 'schedule'
     runs-on: ubuntu-slim
+    environment: ci-protected
     timeout-minutes: 5
     steps:
       - name: Checkout

.github/workflows/post-merge-beta-cherry-pick.yml

@@ -183,6 +183,7 @@ jobs:
       - cherry-pick-to-latest-release
     if: needs.resolve-cherry-pick-request.outputs.should_cherrypick == 'true' && needs.resolve-cherry-pick-request.result == 'success' && needs.cherry-pick-to-latest-release.result == 'success'
     runs-on: ubuntu-slim
+    environment: ci-protected
     timeout-minutes: 10
     steps:
       - name: Checkout
@@ -232,6 +233,7 @@ jobs:
       - cherry-pick-to-latest-release
     if: always() && needs.resolve-cherry-pick-request.outputs.should_cherrypick == 'true' && (needs.resolve-cherry-pick-request.result == 'failure' || needs.cherry-pick-to-latest-release.result == 'failure')
     runs-on: ubuntu-slim
+    environment: ci-protected
     timeout-minutes: 10
     steps:
       - name: Checkout

.github/workflows/pr-desktop-build.yml

@@ -63,7 +63,7 @@ jobs:
           targets: ${{ matrix.target }}
 
       - name: Cache Cargo registry and build
-        uses: actions/cache@cdf6c1fa76f9f475f3d7449005a359c84ca0f306 # zizmor: ignore[cache-poisoning]
+        uses: actions/cache@668228422ae6a00e4ad889ee87cd7109ec5666a7 # zizmor: ignore[cache-poisoning]
         with:
           path: |
             ~/.cargo/bin/

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

@@ -284,7 +284,7 @@ jobs:
 
       - name: Cache playwright cache
         # zizmor: ignore[cache-poisoning] ephemeral runners; no release artifacts
-        uses: runs-on/cache@50350ad4242587b6c8c2baa2e740b1bc11285ff4 # ratchet:runs-on/cache@v4
+        uses: runs-on/cache@a5f51d6f3fece787d03b7b4e981c82538a0654ed # ratchet:runs-on/cache@v4
         with:
           path: ~/.cache/ms-playwright
           key: ${{ runner.os }}-playwright-npm-${{ hashFiles('web/package-lock.json') }}
@@ -626,7 +626,7 @@ jobs:
 
       - name: Cache playwright cache
         # zizmor: ignore[cache-poisoning] ephemeral runners; no release artifacts
-        uses: runs-on/cache@50350ad4242587b6c8c2baa2e740b1bc11285ff4 # ratchet:runs-on/cache@v4
+        uses: runs-on/cache@a5f51d6f3fece787d03b7b4e981c82538a0654ed # ratchet:runs-on/cache@v4
         with:
           path: ~/.cache/ms-playwright
           key: ${{ runner.os }}-playwright-npm-${{ hashFiles('web/package-lock.json') }}

.github/workflows/pr-python-checks.yml

@@ -56,7 +56,7 @@ jobs:
 
       - name: Cache mypy cache
         if: ${{ vars.DISABLE_MYPY_CACHE != 'true' }}
-        uses: runs-on/cache@50350ad4242587b6c8c2baa2e740b1bc11285ff4 # ratchet:runs-on/cache@v4
+        uses: runs-on/cache@a5f51d6f3fece787d03b7b4e981c82538a0654ed # ratchet:runs-on/cache@v4
         with:
           path: .mypy_cache
           key: mypy-${{ runner.os }}-${{ github.base_ref || github.event.merge_group.base_ref || 'main' }}-${{ hashFiles('**/*.py', '**/*.pyi', 'pyproject.toml') }}

.github/workflows/pr-python-model-tests.yml

@@ -31,6 +31,7 @@ jobs:
       - runner=4cpu-linux-arm64
       - "run-id=${{ github.run_id }}-model-check"
       - "extras=ecr-cache"
+    environment: ci-protected
     timeout-minutes: 45
 
     env:

.github/workflows/preview.yml

@@ -15,6 +15,7 @@ permissions:
 jobs:
   Deploy-Preview:
     runs-on: ubuntu-latest
+    environment: ci-protected
     timeout-minutes: 30
     steps:
       - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd

.github/workflows/release-cli.yml

@@ -13,15 +13,6 @@ jobs:
     permissions:
       id-token: write
     timeout-minutes: 10
-    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" }
     steps:
       - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # ratchet:actions/checkout@v6
         with:
@@ -31,9 +22,11 @@ jobs:
           enable-cache: false
           version: "0.9.9"
       - run: |
-          GOOS="${{ matrix.os-arch.goos }}" \
-          GOARCH="${{ matrix.os-arch.goarch }}" \
-          uv build --wheel
+          for goos in linux windows darwin; do
+            for goarch in amd64 arm64; do
+              GOOS="$goos" GOARCH="$goarch" uv build --wheel
+            done
+          done
         working-directory: cli
       - run: uv publish
         working-directory: cli

.github/workflows/storybook-deploy.yml

@@ -25,6 +25,7 @@ permissions:
 jobs:
   Deploy-Storybook:
     runs-on: ubuntu-latest
+    environment: ci-protected
     timeout-minutes: 30
     steps:
       - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # ratchet:actions/checkout@v4
@@ -54,6 +55,7 @@ jobs:
     needs: Deploy-Storybook
     if: always() && needs.Deploy-Storybook.result == 'failure'
     runs-on: ubuntu-latest
+    environment: ci-protected
     timeout-minutes: 10
     steps:
       - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # ratchet:actions/checkout@v4

.github/workflows/sync_foss.yml

@@ -9,6 +9,7 @@ on:
 jobs:
   sync-foss:
     runs-on: ubuntu-latest
+    environment: ci-protected
     timeout-minutes: 45
     permissions:
       contents: read

.github/workflows/tag-nightly.yml

@@ -11,6 +11,7 @@ permissions:
 jobs:
   create-and-push-tag:
     runs-on: ubuntu-slim
+    environment: ci-protected
     timeout-minutes: 45
 
     steps:

.greptile/config.json

@@ -0,0 +1,64 @@
+{
+    "labels": [],
+    "comment": "",
+    "fixWithAI": true,
+    "hideFooter": false,
+    "strictness": 3,
+    "statusCheck": true,
+    "commentTypes": [
+      "logic",
+      "syntax",
+      "style"
+    ],
+    "instructions": "",
+    "disabledLabels": [],
+    "excludeAuthors": [
+      "dependabot[bot]",
+      "renovate[bot]"
+    ],
+    "ignoreKeywords": "",
+    "ignorePatterns": "",
+    "includeAuthors": [],
+    "summarySection": {
+      "included": true,
+      "collapsible": false,
+      "defaultOpen": false
+    },
+    "excludeBranches": [],
+    "fileChangeLimit": 300,
+    "includeBranches": [],
+    "includeKeywords": "",
+    "triggerOnUpdates": true,
+    "updateExistingSummaryComment": true,
+    "updateSummaryOnly": false,
+    "issuesTableSection": {
+      "included": true,
+      "collapsible": false,
+      "defaultOpen": false
+    },
+    "statusCommentsEnabled": true,
+    "confidenceScoreSection": {
+      "included": true,
+      "collapsible": false
+    },
+    "sequenceDiagramSection": {
+      "included": true,
+      "collapsible": false,
+      "defaultOpen": false
+    },
+    "shouldUpdateDescription": false,
+    "rules": [
+      {
+        "scope": ["web/**"],
+        "rule": "In Onyx's Next.js app, the `app/ee/admin/` directory is a filesystem convention for Enterprise Edition route overrides — it does NOT add an `/ee/` prefix to the URL. Both `app/admin/groups/page.tsx` and `app/ee/admin/groups/page.tsx` serve the same URL `/admin/groups`. Hardcoded `/admin/...` paths in router.push() calls are correct and do NOT break EE deployments. Do not flag hardcoded admin paths as bugs."
+      },
+      {
+        "scope": ["web/**"],
+        "rule": "In Onyx, each API key creates a unique user row in the database with a unique `user_id` (UUID). There is a 1:1 mapping between API keys and their backing user records. Multiple API keys do NOT share the same `user_id`. Do not flag potential duplicate row IDs when using `user_id` from API key descriptors."
+      },
+      {
+        "scope": ["backend/**/*.py"],
+        "rule": "Never raise HTTPException directly in business code. Use `raise OnyxError(OnyxErrorCode.XXX, \"message\")` from `onyx.error_handling.exceptions`. A global FastAPI exception handler converts OnyxError into structured JSON responses with {\"error_code\": \"...\", \"detail\": \"...\"}. Error codes are defined in `onyx.error_handling.error_codes.OnyxErrorCode`. For upstream errors with dynamic HTTP status codes, use `status_code_override`: `raise OnyxError(OnyxErrorCode.BAD_GATEWAY, detail, status_code_override=upstream_status)`."
+      }
+    ]
+}

.greptile/files.json

@@ -0,0 +1,57 @@
+[
+  {
+    "scope": [],
+    "path": "contributing_guides/best_practices.md",
+    "description": "Best practices for contributing to the codebase"
+  },
+  {
+    "scope": ["web/**"],
+    "path": "web/AGENTS.md",
+    "description": "Frontend coding standards for the web directory"
+  },
+  {
+    "scope": ["web/**"],
+    "path": "web/tests/README.md",
+    "description": "Frontend testing guide and conventions"
+  },
+  {
+    "scope": ["web/**"],
+    "path": "web/CLAUDE.md",
+    "description": "Single source of truth for frontend coding standards"
+  },
+  {
+    "scope": ["web/**"],
+    "path": "web/lib/opal/README.md",
+    "description": "Opal component library usage guide"
+  },
+  {
+    "scope": ["backend/**"],
+    "path": "backend/tests/README.md",
+    "description": "Backend testing guide covering all 4 test types, fixtures, and conventions"
+  },
+  {
+    "scope": ["backend/onyx/connectors/**"],
+    "path": "backend/onyx/connectors/README.md",
+    "description": "Connector development guide covering design, interfaces, and required changes"
+  },
+  {
+    "scope": [],
+    "path": "CLAUDE.md",
+    "description": "Project instructions and coding standards"
+  },
+  {
+    "scope": [],
+    "path": "backend/alembic/README.md",
+    "description": "Migration guidance, including multi-tenant migration behavior"
+  },
+  {
+    "scope": [],
+    "path": "deployment/helm/charts/onyx/values-lite.yaml",
+    "description": "Lite deployment Helm values and service assumptions"
+  },
+  {
+    "scope": [],
+    "path": "deployment/docker_compose/docker-compose.onyx-lite.yml",
+    "description": "Lite deployment Docker Compose overlay and disabled service behavior"
+  }
+]

.greptile/rules.md

@@ -0,0 +1,39 @@
+# Greptile Review Rules
+
+## Type Annotations
+
+Use explicit type annotations for variables to enhance code clarity, especially when moving type hints around in the code.
+
+## Best Practices
+
+Use `contributing_guides/best_practices.md` as core review context. Prefer consistency with existing patterns, fix issues in code you touch, avoid tacking new features onto muddy interfaces, fail loudly instead of silently swallowing errors, keep code strictly typed, preserve clear state boundaries, remove duplicate or dead logic, break up overly long functions, avoid hidden import-time side effects, respect module boundaries, and favor correctness-by-construction over relying on callers to use an API correctly.
+
+## TODOs
+
+Whenever a TODO is added, there must always be an associated name or ticket with that TODO in the style of `TODO(name): ...` or `TODO(1234): ...`
+
+## Debugging Code
+
+Remove temporary debugging code before merging to production, especially tenant-specific debugging logs.
+
+## Hardcoded Booleans
+
+When hardcoding a boolean variable to a constant value, remove the variable entirely and clean up all places where it's used rather than just setting it to a constant.
+
+## Multi-tenant vs Single-tenant
+
+Code changes must consider both multi-tenant and single-tenant deployments. In multi-tenant mode, preserve tenant isolation, ensure tenant context is propagated correctly, and avoid assumptions that only hold for a single shared schema or globally shared state. In single-tenant mode, avoid introducing unnecessary tenant-specific requirements or cloud-only control-plane dependencies.
+
+## Nginx Routing — New Backend Routes
+
+Whenever a new backend route is added that does NOT start with `/api`, it must also be explicitly added to ALL nginx configs:
+- `deployment/helm/charts/onyx/templates/nginx-conf.yaml` (Helm/k8s)
+- `deployment/data/nginx/app.conf.template` (docker-compose dev)
+- `deployment/data/nginx/app.conf.template.prod` (docker-compose prod)
+- `deployment/data/nginx/app.conf.template.no-letsencrypt` (docker-compose no-letsencrypt)
+
+Routes not starting with `/api` are not caught by the existing `^/(api|openapi\.json)` location block and will fall through to `location /`, which proxies to the Next.js web server and returns an HTML 404. The new location block must be placed before the `/api` block. Examples of routes that need this treatment: `/scim`, `/mcp`.
+
+## Full vs Lite Deployments
+
+Code changes must consider both regular Onyx deployments and Onyx lite deployments. Lite deployments disable the vector DB, Redis, model servers, and background workers by default, use PostgreSQL-backed cache/auth/file storage, and rely on the API server to handle background work. Do not assume those services are available unless the code path is explicitly limited to full deployments.

.pre-commit-config.yaml

@@ -122,7 +122,7 @@ repos:
     rev: 5d1e709b7be35cb2025444e19de266b056b7b7ee # frozen: v2.10.1
     hooks:
       - id: golangci-lint
-        language_version: "1.26.0"
+        language_version: "1.26.1"
         entry: bash -c "find . -name go.mod -not -path './.venv/*' -print0 | xargs -0 -I{} bash -c 'cd \"$(dirname {})\" && golangci-lint run ./...'"
 
   - repo: https://github.com/astral-sh/ruff-pre-commit

README.md

@@ -35,7 +35,7 @@ Onyx comes loaded with advanced features like Agents, Web Search, RAG, MCP, Deep
 > [!TIP]
 > Run Onyx with one command (or see deployment section below):
 > ```
-> curl -fsSL https://raw.githubusercontent.com/onyx-dot-app/onyx/main/deployment/docker_compose/install.sh > install.sh && chmod +x install.sh && ./install.sh
+> curl -fsSL https://onyx.app/install_onyx.sh | bash
 > ```
 
 ****

backend/alembic/versions/1d78c0ca7853_remove_voice_provider_deleted_column.py

@@ -0,0 +1,35 @@
+"""remove voice_provider deleted column
+
+Revision ID: 1d78c0ca7853
+Revises: a3f8b2c1d4e5
+Create Date: 2026-03-26 11:30:53.883127
+
+"""
+
+from alembic import op
+import sqlalchemy as sa
+
+
+# revision identifiers, used by Alembic.
+revision = "1d78c0ca7853"
+down_revision = "a3f8b2c1d4e5"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    # Hard-delete any soft-deleted rows before dropping the column
+    op.execute("DELETE FROM voice_provider WHERE deleted = true")
+    op.drop_column("voice_provider", "deleted")
+
+
+def downgrade() -> None:
+    op.add_column(
+        "voice_provider",
+        sa.Column(
+            "deleted",
+            sa.Boolean(),
+            nullable=False,
+            server_default=sa.text("false"),
+        ),
+    )

backend/ee/onyx/background/celery/tasks/doc_permission_syncing/tasks.py

@@ -28,6 +28,7 @@ from onyx.access.models import DocExternalAccess
 from onyx.access.models import ElementExternalAccess
 from onyx.background.celery.apps.app_base import task_logger
 from onyx.background.celery.celery_redis import celery_find_task
+from onyx.background.celery.celery_redis import celery_get_broker_client
 from onyx.background.celery.celery_redis import celery_get_queue_length
 from onyx.background.celery.celery_redis import celery_get_queued_task_ids
 from onyx.background.celery.celery_redis import celery_get_unacked_task_ids
@@ -187,7 +188,6 @@ def check_for_doc_permissions_sync(self: Task, *, tenant_id: str) -> bool | None
     # (which lives on a different db number)
     r = get_redis_client()
     r_replica = get_redis_replica_client()
-    r_celery: Redis = self.app.broker_connection().channel().client  # type: ignore
 
     lock_beat: RedisLock = r.lock(
         OnyxRedisLocks.CHECK_CONNECTOR_DOC_PERMISSIONS_SYNC_BEAT_LOCK,
@@ -227,6 +227,7 @@ def check_for_doc_permissions_sync(self: Task, *, tenant_id: str) -> bool | None
             # tasks can be in the queue in redis, in reserved tasks (prefetched by the worker),
             # or be currently executing
             try:
+                r_celery = celery_get_broker_client(self.app)
                 validate_permission_sync_fences(
                     tenant_id, r, r_replica, r_celery, lock_beat
                 )
@@ -473,6 +474,8 @@ def connector_permission_sync_generator_task(
             cc_pair = get_connector_credential_pair_from_id(
                 db_session=db_session,
                 cc_pair_id=cc_pair_id,
+                eager_load_connector=True,
+                eager_load_credential=True,
             )
             if cc_pair is None:
                 raise ValueError(

backend/ee/onyx/background/celery/tasks/external_group_syncing/tasks.py

@@ -29,6 +29,7 @@ from ee.onyx.external_permissions.sync_params import (
 from ee.onyx.external_permissions.sync_params import get_source_perm_sync_config
 from onyx.background.celery.apps.app_base import task_logger
 from onyx.background.celery.celery_redis import celery_find_task
+from onyx.background.celery.celery_redis import celery_get_broker_client
 from onyx.background.celery.celery_redis import celery_get_unacked_task_ids
 from onyx.background.celery.tasks.beat_schedule import CLOUD_BEAT_MULTIPLIER_DEFAULT
 from onyx.background.error_logging import emit_background_error
@@ -162,7 +163,6 @@ def check_for_external_group_sync(self: Task, *, tenant_id: str) -> bool | None:
     # (which lives on a different db number)
     r = get_redis_client()
     r_replica = get_redis_replica_client()
-    r_celery: Redis = self.app.broker_connection().channel().client  # type: ignore
 
     lock_beat: RedisLock = r.lock(
         OnyxRedisLocks.CHECK_CONNECTOR_EXTERNAL_GROUP_SYNC_BEAT_LOCK,
@@ -221,6 +221,7 @@ def check_for_external_group_sync(self: Task, *, tenant_id: str) -> bool | None:
             # tasks can be in the queue in redis, in reserved tasks (prefetched by the worker),
             # or be currently executing
             try:
+                r_celery = celery_get_broker_client(self.app)
                 validate_external_group_sync_fences(
                     tenant_id, self.app, r, r_replica, r_celery, lock_beat
                 )

backend/ee/onyx/external_permissions/slack/doc_sync.py

@@ -8,6 +8,7 @@ from ee.onyx.external_permissions.slack.utils import fetch_user_id_to_email_map
 from onyx.access.models import DocExternalAccess
 from onyx.access.models import ExternalAccess
 from onyx.connectors.credentials_provider import OnyxDBCredentialsProvider
+from onyx.connectors.interfaces import SecondsSinceUnixEpoch
 from onyx.connectors.models import HierarchyNode
 from onyx.connectors.slack.connector import get_channels
 from onyx.connectors.slack.connector import make_paginated_slack_api_call
@@ -105,9 +106,11 @@ def _get_slack_document_access(
     slack_connector: SlackConnector,
     channel_permissions: dict[str, ExternalAccess],  # noqa: ARG001
     callback: IndexingHeartbeatInterface | None,
+    indexing_start: SecondsSinceUnixEpoch | None = None,
 ) -> Generator[DocExternalAccess, None, None]:
     slim_doc_generator = slack_connector.retrieve_all_slim_docs_perm_sync(
-        callback=callback
+        callback=callback,
+        start=indexing_start,
     )
 
     for doc_metadata_batch in slim_doc_generator:
@@ -180,9 +183,15 @@ def slack_doc_sync(
 
     slack_connector = SlackConnector(**cc_pair.connector.connector_specific_config)
     slack_connector.set_credentials_provider(provider)
+    indexing_start_ts: SecondsSinceUnixEpoch | None = (
+        cc_pair.connector.indexing_start.timestamp()
+        if cc_pair.connector.indexing_start is not None
+        else None
+    )
 
     yield from _get_slack_document_access(
-        slack_connector,
+        slack_connector=slack_connector,
         channel_permissions=channel_permissions,
         callback=callback,
+        indexing_start=indexing_start_ts,
     )

backend/ee/onyx/external_permissions/utils.py

@@ -6,6 +6,7 @@ from onyx.access.models import ElementExternalAccess
 from onyx.access.models import ExternalAccess
 from onyx.access.models import NodeExternalAccess
 from onyx.configs.constants import DocumentSource
+from onyx.connectors.interfaces import SecondsSinceUnixEpoch
 from onyx.connectors.interfaces import SlimConnectorWithPermSync
 from onyx.connectors.models import HierarchyNode
 from onyx.db.models import ConnectorCredentialPair
@@ -40,10 +41,19 @@ def generic_doc_sync(
 
     logger.info(f"Starting {doc_source} doc sync for CC Pair ID: {cc_pair.id}")
 
+    indexing_start: SecondsSinceUnixEpoch | None = (
+        cc_pair.connector.indexing_start.timestamp()
+        if cc_pair.connector.indexing_start is not None
+        else None
+    )
+
     newly_fetched_doc_ids: set[str] = set()
 
     logger.info(f"Fetching all slim documents from {doc_source}")
-    for doc_batch in slim_connector.retrieve_all_slim_docs_perm_sync(callback=callback):
+    for doc_batch in slim_connector.retrieve_all_slim_docs_perm_sync(
+        start=indexing_start,
+        callback=callback,
+    ):
         logger.info(f"Got {len(doc_batch)} slim documents from {doc_source}")
 
         if callback:

backend/onyx/background/celery/celery_redis.py

@@ -1,5 +1,6 @@
 # These are helper objects for tracking the keys we need to write in redis
 import json
+import threading
 from typing import Any
 from typing import cast
 
@@ -7,7 +8,59 @@ from celery import Celery
 from redis import Redis
 
 from onyx.background.celery.configs.base import CELERY_SEPARATOR
+from onyx.configs.app_configs import REDIS_HEALTH_CHECK_INTERVAL
 from onyx.configs.constants import OnyxCeleryPriority
+from onyx.configs.constants import REDIS_SOCKET_KEEPALIVE_OPTIONS
+
+
+_broker_client: Redis | None = None
+_broker_url: str | None = None
+_broker_client_lock = threading.Lock()
+
+
+def celery_get_broker_client(app: Celery) -> Redis:
+    """Return a shared Redis client connected to the Celery broker DB.
+
+    Uses a module-level singleton so all tasks on a worker share one
+    connection instead of creating a new one per call. The client
+    connects directly to the broker Redis DB (parsed from the broker URL).
+
+    Thread-safe via lock — safe for use in Celery thread-pool workers.
+
+    Usage:
+        r_celery = celery_get_broker_client(self.app)
+        length = celery_get_queue_length(queue, r_celery)
+    """
+    global _broker_client, _broker_url
+    with _broker_client_lock:
+        url = app.conf.broker_url
+        if _broker_client is not None and _broker_url == url:
+            try:
+                _broker_client.ping()
+                return _broker_client
+            except Exception:
+                try:
+                    _broker_client.close()
+                except Exception:
+                    pass
+                _broker_client = None
+        elif _broker_client is not None:
+            try:
+                _broker_client.close()
+            except Exception:
+                pass
+            _broker_client = None
+
+        _broker_url = url
+        _broker_client = Redis.from_url(
+            url,
+            decode_responses=False,
+            health_check_interval=REDIS_HEALTH_CHECK_INTERVAL,
+            socket_keepalive=True,
+            socket_keepalive_options=REDIS_SOCKET_KEEPALIVE_OPTIONS,
+            retry_on_timeout=True,
+        )
+        return _broker_client
 
 
 def celery_get_unacked_length(r: Redis) -> int:

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

@@ -14,6 +14,7 @@ from redis.lock import Lock as RedisLock
 from sqlalchemy.orm import Session
 
 from onyx.background.celery.apps.app_base import task_logger
+from onyx.background.celery.celery_redis import celery_get_broker_client
 from onyx.background.celery.celery_redis import celery_get_queue_length
 from onyx.background.celery.celery_redis import celery_get_queued_task_ids
 from onyx.configs.app_configs import JOB_TIMEOUT
@@ -132,7 +133,6 @@ def revoke_tasks_blocking_deletion(
 def check_for_connector_deletion_task(self: Task, *, tenant_id: str) -> bool | None:
     r = get_redis_client()
     r_replica = get_redis_replica_client()
-    r_celery: Redis = self.app.broker_connection().channel().client  # type: ignore
 
     lock_beat: RedisLock = r.lock(
         OnyxRedisLocks.CHECK_CONNECTOR_DELETION_BEAT_LOCK,
@@ -149,6 +149,7 @@ def check_for_connector_deletion_task(self: Task, *, tenant_id: str) -> bool | N
         if not r.exists(OnyxRedisSignals.BLOCK_VALIDATE_CONNECTOR_DELETION_FENCES):
             # clear fences that don't have associated celery tasks in progress
             try:
+                r_celery = celery_get_broker_client(self.app)
                 validate_connector_deletion_fences(
                     tenant_id, r, r_replica, r_celery, lock_beat
                 )

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

@@ -22,6 +22,7 @@ from sqlalchemy.orm import Session
 
 from onyx.background.celery.apps.app_base import task_logger
 from onyx.background.celery.celery_redis import celery_find_task
+from onyx.background.celery.celery_redis import celery_get_broker_client
 from onyx.background.celery.celery_redis import celery_get_unacked_task_ids
 from onyx.background.celery.celery_utils import httpx_init_vespa_pool
 from onyx.background.celery.memory_monitoring import emit_process_memory
@@ -449,7 +450,7 @@ def check_indexing_completion(
             ):
                 # Check if the task exists in the celery queue
                 # This handles the case where Redis dies after task creation but before task execution
-                redis_celery = task.app.broker_connection().channel().client  # type: ignore
+                redis_celery = celery_get_broker_client(task.app)
                 task_exists = celery_find_task(
                     attempt.celery_task_id,
                     OnyxCeleryQueues.CONNECTOR_DOC_FETCHING,

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

@@ -1,6 +1,5 @@
 import json
 import time
-from collections.abc import Callable
 from datetime import timedelta
 from itertools import islice
 from typing import Any
@@ -19,6 +18,7 @@ from sqlalchemy import text
 from sqlalchemy.orm import Session
 
 from onyx.background.celery.apps.app_base import task_logger
+from onyx.background.celery.celery_redis import celery_get_broker_client
 from onyx.background.celery.celery_redis import celery_get_queue_length
 from onyx.background.celery.celery_redis import celery_get_unacked_task_ids
 from onyx.background.celery.memory_monitoring import emit_process_memory
@@ -698,31 +698,27 @@ def monitor_background_processes(self: Task, *, tenant_id: str) -> None:
         return None
 
     try:
-        # Get Redis client for Celery broker
-        redis_celery = self.app.broker_connection().channel().client  # type: ignore
         redis_std = get_redis_client()
 
-        # Define metric collection functions and their dependencies
-        metric_functions: list[Callable[[], list[Metric]]] = [
-            lambda: _collect_queue_metrics(redis_celery),
-            lambda: _collect_connector_metrics(db_session, redis_std),
-            lambda: _collect_sync_metrics(db_session, redis_std),
-        ]
+        # Collect queue metrics with broker connection
+        r_celery = celery_get_broker_client(self.app)
+        queue_metrics = _collect_queue_metrics(r_celery)
 
-        # Collect and log each metric
+        # Collect remaining metrics (no broker connection needed)
         with get_session_with_current_tenant() as db_session:
-            for metric_fn in metric_functions:
-                metrics = metric_fn()
-                for metric in metrics:
-                    # double check to make sure we aren't double-emitting metrics
-                    if metric.key is None or not _has_metric_been_emitted(
-                        redis_std, metric.key
-                    ):
-                        metric.log()
-                        metric.emit(tenant_id)
+            all_metrics: list[Metric] = queue_metrics
+            all_metrics.extend(_collect_connector_metrics(db_session, redis_std))
+            all_metrics.extend(_collect_sync_metrics(db_session, redis_std))
 
-                    if metric.key is not None:
-                        _mark_metric_as_emitted(redis_std, metric.key)
+            for metric in all_metrics:
+                if metric.key is None or not _has_metric_been_emitted(
+                    redis_std, metric.key
+                ):
+                    metric.log()
+                    metric.emit(tenant_id)
+
+                if metric.key is not None:
+                    _mark_metric_as_emitted(redis_std, metric.key)
 
         task_logger.info("Successfully collected background metrics")
     except SoftTimeLimitExceeded:
@@ -890,7 +886,7 @@ def monitor_celery_queues_helper(
 ) -> None:
     """A task to monitor all celery queue lengths."""
 
-    r_celery = task.app.broker_connection().channel().client  # type: ignore
+    r_celery = celery_get_broker_client(task.app)
     n_celery = celery_get_queue_length(OnyxCeleryQueues.PRIMARY, r_celery)
     n_docfetching = celery_get_queue_length(
         OnyxCeleryQueues.CONNECTOR_DOC_FETCHING, r_celery
@@ -1080,7 +1076,7 @@ def cloud_monitor_celery_pidbox(
     num_deleted = 0
 
     MAX_PIDBOX_IDLE = 24 * 3600  # 1 day in seconds
-    r_celery: Redis = self.app.broker_connection().channel().client  # type: ignore
+    r_celery = celery_get_broker_client(self.app)
     for key in r_celery.scan_iter("*.reply.celery.pidbox"):
         key_bytes = cast(bytes, key)
         key_str = key_bytes.decode("utf-8")

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

@@ -17,6 +17,7 @@ from sqlalchemy.orm import Session
 
 from onyx.background.celery.apps.app_base import task_logger
 from onyx.background.celery.celery_redis import celery_find_task
+from onyx.background.celery.celery_redis import celery_get_broker_client
 from onyx.background.celery.celery_redis import celery_get_queue_length
 from onyx.background.celery.celery_redis import celery_get_queued_task_ids
 from onyx.background.celery.celery_redis import celery_get_unacked_task_ids
@@ -203,7 +204,6 @@ def _is_pruning_due(cc_pair: ConnectorCredentialPair) -> bool:
 def check_for_pruning(self: Task, *, tenant_id: str) -> bool | None:
     r = get_redis_client()
     r_replica = get_redis_replica_client()
-    r_celery: Redis = self.app.broker_connection().channel().client  # type: ignore
 
     lock_beat: RedisLock = r.lock(
         OnyxRedisLocks.CHECK_PRUNE_BEAT_LOCK,
@@ -261,6 +261,7 @@ def check_for_pruning(self: Task, *, tenant_id: str) -> bool | None:
             # tasks can be in the queue in redis, in reserved tasks (prefetched by the worker),
             # or be currently executing
             try:
+                r_celery = celery_get_broker_client(self.app)
                 validate_pruning_fences(tenant_id, r, r_replica, r_celery, lock_beat)
             except Exception:
                 task_logger.exception("Exception while validating pruning fences")

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

@@ -16,6 +16,7 @@ from sqlalchemy.orm import Session
 
 from onyx.access.access import build_access_for_user_files
 from onyx.background.celery.apps.app_base import task_logger
+from onyx.background.celery.celery_redis import celery_get_broker_client
 from onyx.background.celery.celery_redis import celery_get_queue_length
 from onyx.background.celery.celery_utils import httpx_init_vespa_pool
 from onyx.background.celery.tasks.shared.RetryDocumentIndex import RetryDocumentIndex
@@ -105,7 +106,7 @@ def _user_file_delete_queued_key(user_file_id: str | UUID) -> str:
 
 
 def get_user_file_project_sync_queue_depth(celery_app: Celery) -> int:
-    redis_celery: Redis = celery_app.broker_connection().channel().client  # type: ignore
+    redis_celery = celery_get_broker_client(celery_app)
     return celery_get_queue_length(
         OnyxCeleryQueues.USER_FILE_PROJECT_SYNC, redis_celery
     )
@@ -238,7 +239,7 @@ def check_user_file_processing(self: Task, *, tenant_id: str) -> None:
     skipped_guard = 0
     try:
         # --- Protection 1: queue depth backpressure ---
-        r_celery = self.app.broker_connection().channel().client  # type: ignore
+        r_celery = celery_get_broker_client(self.app)
         queue_len = celery_get_queue_length(
             OnyxCeleryQueues.USER_FILE_PROCESSING, r_celery
         )
@@ -591,7 +592,7 @@ def check_for_user_file_delete(self: Task, *, tenant_id: str) -> None:
         # --- Protection 1: queue depth backpressure ---
         # NOTE: must use the broker's Redis client (not redis_client) because
         # Celery queues live on a separate Redis DB with CELERY_SEPARATOR keys.
-        r_celery: Redis = self.app.broker_connection().channel().client  # type: ignore
+        r_celery = celery_get_broker_client(self.app)
         queue_len = celery_get_queue_length(OnyxCeleryQueues.USER_FILE_DELETE, r_celery)
         if queue_len > USER_FILE_DELETE_MAX_QUEUE_DEPTH:
             task_logger.warning(

backend/onyx/chat/chat_state.py

@@ -1,19 +1,8 @@
 import threading
-import time
-from collections.abc import Callable
-from collections.abc import Generator
-from queue import Empty
 
 from onyx.chat.citation_processor import CitationMapping
-from onyx.chat.emitter import Emitter
 from onyx.context.search.models import SearchDoc
-from onyx.server.query_and_chat.placement import Placement
-from onyx.server.query_and_chat.streaming_models import OverallStop
-from onyx.server.query_and_chat.streaming_models import Packet
-from onyx.server.query_and_chat.streaming_models import PacketException
 from onyx.tools.models import ToolCallInfo
-from onyx.utils.threadpool_concurrency import run_in_background
-from onyx.utils.threadpool_concurrency import wait_on_background
 
 # Type alias for search doc deduplication key
 # Simple key: just document_id (str)
@@ -159,114 +148,3 @@ class ChatStateContainer:
         """Thread-safe getter for emitted citations (returns a copy)."""
         with self._lock:
             return self._emitted_citations.copy()
-
-
-def run_chat_loop_with_state_containers(
-    chat_loop_func: Callable[[Emitter, ChatStateContainer], None],
-    completion_callback: Callable[[ChatStateContainer], None],
-    is_connected: Callable[[], bool],
-    emitter: Emitter,
-    state_container: ChatStateContainer,
-) -> Generator[Packet, None]:
-    """
-    Explicit wrapper function that runs a function in a background thread
-    with event streaming capabilities.
-
-    The wrapped function should accept emitter as first arg and use it to emit
-    Packet objects. This wrapper polls every 300ms to check if stop signal is set.
-
-    Args:
-        func: The function to wrap (should accept emitter and state_container as first and second args)
-        completion_callback: Callback function to call when the function completes
-        emitter: Emitter instance for sending packets
-        state_container: ChatStateContainer instance for accumulating state
-        is_connected: Callable that returns False when stop signal is set
-
-    Usage:
-        packets = run_chat_loop_with_state_containers(
-            my_func,
-            completion_callback=completion_callback,
-            emitter=emitter,
-            state_container=state_container,
-            is_connected=check_func,
-        )
-        for packet in packets:
-            # Process packets
-            pass
-    """
-
-    def run_with_exception_capture() -> None:
-        try:
-            chat_loop_func(emitter, state_container)
-        except Exception as e:
-            # If execution fails, emit an exception packet
-            emitter.emit(
-                Packet(
-                    placement=Placement(turn_index=0),
-                    obj=PacketException(type="error", exception=e),
-                )
-            )
-
-    # Run the function in a background thread
-    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
-            # the 300ms timeout is to avoid busy-waiting and to allow the stop signal to be checked regularly
-            try:
-                pkt = emitter.bus.get(timeout=0.3)
-            except Empty:
-                if not is_connected():
-                    # 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:
-                # 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.
-        if is_connected():
-            wait_on_background(thread)
-        try:
-            completion_callback(state_container)
-        except Exception as e:
-            emitter.emit(
-                Packet(
-                    placement=Placement(turn_index=last_turn_index + 1),
-                    obj=PacketException(type="error", exception=e),
-                )
-            )

backend/onyx/chat/emitter.py

@@ -1,19 +1,45 @@
+import logging
+import queue
 from queue import Queue
+from typing import Any
 
+from onyx.server.query_and_chat.placement import Placement
 from onyx.server.query_and_chat.streaming_models import Packet
 
+logger = logging.getLogger(__name__)
+
 
 class Emitter:
-    """Use this inside tools to emit arbitrary UI progress."""
+    """Routes packets from LLM/tool execution to the ``_run_models`` drain loop.
 
-    def __init__(self, bus: Queue):
-        self.bus = bus
+    Tags every packet with ``model_index`` and places it on ``merged_queue``
+    as a ``(model_idx, packet)`` tuple for ordered consumption downstream.
+
+    Args:
+        merged_queue: Shared queue owned by ``_run_models``.
+        model_idx: Index embedded in packet placements (``0`` for N=1 runs).
+    """
+
+    def __init__(
+        self,
+        merged_queue: Queue[Any],
+        model_idx: int = 0,
+    ) -> None:
+        self._model_idx = model_idx
+        self._merged_queue = merged_queue
 
     def emit(self, packet: Packet) -> None:
-        self.bus.put(packet)  # Thread-safe
-
-
-def get_default_emitter() -> Emitter:
-    bus: Queue[Packet] = Queue()
-    emitter = Emitter(bus)
-    return emitter
+        base = packet.placement or Placement(turn_index=0)
+        tagged = Packet(
+            placement=base.model_copy(update={"model_index": self._model_idx}),
+            obj=packet.obj,
+        )
+        try:
+            self._merged_queue.put((self._model_idx, tagged), timeout=3.0)
+        except queue.Full:
+            # Drain loop is gone (e.g. GeneratorExit on disconnect); discard packet.
+            logger.warning(
+                "Emitter model_idx=%d: queue full after 3s timeout, dropping packet %s",
+                self._model_idx,
+                type(packet.obj).__name__,
+            )

backend/onyx/configs/app_configs.py

@@ -44,6 +44,31 @@ SEND_USER_METADATA_TO_LLM_PROVIDER = (
 # User Facing Features Configs
 #####
 BLURB_SIZE = 128  # Number Encoder Tokens included in the chunk blurb
+
+# Hard ceiling for the admin-configurable file upload size (in MB).
+# Self-hosted customers can raise or lower this via the environment variable.
+_raw_max_upload_size_mb = int(os.environ.get("MAX_ALLOWED_UPLOAD_SIZE_MB", "250"))
+if _raw_max_upload_size_mb < 0:
+    logger.warning(
+        "MAX_ALLOWED_UPLOAD_SIZE_MB=%d is negative; falling back to 250",
+        _raw_max_upload_size_mb,
+    )
+    _raw_max_upload_size_mb = 250
+MAX_ALLOWED_UPLOAD_SIZE_MB = _raw_max_upload_size_mb
+
+# Default fallback for the per-user file upload size limit (in MB) when no
+# admin-configured value exists.  Clamped to MAX_ALLOWED_UPLOAD_SIZE_MB at
+# runtime so this never silently exceeds the hard ceiling.
+_raw_default_upload_size_mb = int(
+    os.environ.get("DEFAULT_USER_FILE_MAX_UPLOAD_SIZE_MB", "100")
+)
+if _raw_default_upload_size_mb < 0:
+    logger.warning(
+        "DEFAULT_USER_FILE_MAX_UPLOAD_SIZE_MB=%d is negative; falling back to 100",
+        _raw_default_upload_size_mb,
+    )
+    _raw_default_upload_size_mb = 100
+DEFAULT_USER_FILE_MAX_UPLOAD_SIZE_MB = _raw_default_upload_size_mb
 GENERATIVE_MODEL_ACCESS_CHECK_FREQ = int(
     os.environ.get("GENERATIVE_MODEL_ACCESS_CHECK_FREQ") or 86400
 )  # 1 day
@@ -61,17 +86,6 @@ CACHE_BACKEND = CacheBackendType(
     os.environ.get("CACHE_BACKEND", CacheBackendType.REDIS)
 )
 
-# Maximum token count for a single uploaded file. Files exceeding this are rejected.
-# Defaults to 100k tokens (or 10M when vector DB is disabled).
-_DEFAULT_FILE_TOKEN_LIMIT = 10_000_000 if DISABLE_VECTOR_DB else 100_000
-FILE_TOKEN_COUNT_THRESHOLD = int(
-    os.environ.get("FILE_TOKEN_COUNT_THRESHOLD", str(_DEFAULT_FILE_TOKEN_LIMIT))
-)
-
-# Maximum upload size for a single user file (chat/projects) in MB.
-USER_FILE_MAX_UPLOAD_SIZE_MB = int(os.environ.get("USER_FILE_MAX_UPLOAD_SIZE_MB") or 50)
-USER_FILE_MAX_UPLOAD_SIZE_BYTES = USER_FILE_MAX_UPLOAD_SIZE_MB * 1024 * 1024
-
 # If set to true, will show extra/uncommon connectors in the "Other" category
 SHOW_EXTRA_CONNECTORS = os.environ.get("SHOW_EXTRA_CONNECTORS", "").lower() == "true"
 
@@ -791,6 +805,10 @@ MINI_CHUNK_SIZE = 150
 # This is the number of regular chunks per large chunk
 LARGE_CHUNK_RATIO = 4
 
+# The maximum number of chunks that can be held for 1 document processing batch
+# The purpose of this is to set an upper bound on memory usage
+MAX_CHUNKS_PER_DOC_BATCH = int(os.environ.get("MAX_CHUNKS_PER_DOC_BATCH") or 1000)
+
 # 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"

backend/onyx/connectors/canvas/client.py

@@ -0,0 +1,192 @@
+from __future__ import annotations
+
+import logging
+import re
+from typing import Any
+from urllib.parse import urlparse
+
+from onyx.connectors.cross_connector_utils.rate_limit_wrapper import (
+    rl_requests,
+)
+from onyx.error_handling.error_codes import OnyxErrorCode
+from onyx.error_handling.exceptions import OnyxError
+
+logger = logging.getLogger(__name__)
+
+# Requests timeout in seconds.
+_CANVAS_CALL_TIMEOUT: int = 30
+_CANVAS_API_VERSION: str = "/api/v1"
+# Matches the "next" URL in a Canvas Link header, e.g.:
+#   <https://canvas.example.com/api/v1/courses?page=2>; rel="next"
+# Captures the URL inside the angle brackets.
+_NEXT_LINK_PATTERN: re.Pattern[str] = re.compile(r'<([^>]+)>;\s*rel="next"')
+
+
+_STATUS_TO_ERROR_CODE: dict[int, OnyxErrorCode] = {
+    401: OnyxErrorCode.CREDENTIAL_EXPIRED,
+    403: OnyxErrorCode.INSUFFICIENT_PERMISSIONS,
+    404: OnyxErrorCode.BAD_GATEWAY,
+    429: OnyxErrorCode.RATE_LIMITED,
+}
+
+
+def _error_code_for_status(status_code: int) -> OnyxErrorCode:
+    """Map an HTTP status code to the appropriate OnyxErrorCode.
+
+    Expects a >= 400 status code. Known codes (401, 403, 404, 429) are
+    mapped to specific error codes; all other codes (unrecognised 4xx
+    and 5xx) map to BAD_GATEWAY as unexpected upstream errors.
+    """
+    if status_code in _STATUS_TO_ERROR_CODE:
+        return _STATUS_TO_ERROR_CODE[status_code]
+    return OnyxErrorCode.BAD_GATEWAY
+
+
+class CanvasApiClient:
+    def __init__(
+        self,
+        bearer_token: str,
+        canvas_base_url: str,
+    ) -> None:
+        parsed_base = urlparse(canvas_base_url)
+        if not parsed_base.hostname:
+            raise ValueError("canvas_base_url must include a valid host")
+        if parsed_base.scheme != "https":
+            raise ValueError("canvas_base_url must use https")
+
+        self._bearer_token = bearer_token
+        self.base_url = (
+            canvas_base_url.rstrip("/").removesuffix(_CANVAS_API_VERSION)
+            + _CANVAS_API_VERSION
+        )
+        # Hostname is already validated above; reuse parsed_base instead
+        # of re-parsing.  Used by _parse_next_link to validate pagination URLs.
+        self._expected_host: str = parsed_base.hostname
+
+    def get(
+        self,
+        endpoint: str = "",
+        params: dict[str, Any] | None = None,
+        full_url: str | None = None,
+    ) -> tuple[Any, str | None]:
+        """Make a GET request to the Canvas API.
+
+        Returns a tuple of (json_body, next_url).
+        next_url is parsed from the Link header and is None if there are no more pages.
+        If full_url is provided, it is used directly (for following pagination links).
+
+        Security note: full_url must only be set to values returned by
+        ``_parse_next_link``, which validates the host against the configured
+        Canvas base URL.  Passing an arbitrary URL would leak the bearer token.
+        """
+        # full_url is used when following pagination (Canvas returns the
+        # next-page URL in the Link header).  For the first request we build
+        # the URL from the endpoint name instead.
+        url = full_url if full_url else self._build_url(endpoint)
+        headers = self._build_headers()
+
+        response = rl_requests.get(
+            url,
+            headers=headers,
+            params=params if not full_url else None,
+            timeout=_CANVAS_CALL_TIMEOUT,
+        )
+
+        try:
+            response_json = response.json()
+        except ValueError as e:
+            if response.status_code < 300:
+                raise OnyxError(
+                    OnyxErrorCode.BAD_GATEWAY,
+                    detail=f"Invalid JSON in Canvas response: {e}",
+                )
+            logger.warning(
+                "Failed to parse JSON from Canvas error response (status=%d): %s",
+                response.status_code,
+                e,
+            )
+            response_json = {}
+
+        if response.status_code >= 400:
+            # Try to extract the most specific error message from the
+            # Canvas response body.  Canvas uses three different shapes
+            # depending on the endpoint and error type:
+            default_error: str = response.reason or f"HTTP {response.status_code}"
+            error = default_error
+            if isinstance(response_json, dict):
+                # Shape 1: {"error": {"message": "Not authorized"}}
+                error_field = response_json.get("error")
+                if isinstance(error_field, dict):
+                    response_error = error_field.get("message", "")
+                    if response_error:
+                        error = response_error
+                # Shape 2: {"error": "Invalid access token"}
+                elif isinstance(error_field, str):
+                    error = error_field
+                # Shape 3: {"errors": [{"message": "..."}]}
+                # Used for validation errors.  Only use as fallback if
+                # we didn't already find a more specific message above.
+                if error == default_error:
+                    errors_list = response_json.get("errors")
+                    if isinstance(errors_list, list) and errors_list:
+                        first_error = errors_list[0]
+                        if isinstance(first_error, dict):
+                            msg = first_error.get("message", "")
+                            if msg:
+                                error = msg
+            raise OnyxError(
+                _error_code_for_status(response.status_code),
+                detail=error,
+                status_code_override=response.status_code,
+            )
+
+        next_url = self._parse_next_link(response.headers.get("Link", ""))
+        return response_json, next_url
+
+    def _parse_next_link(self, link_header: str) -> str | None:
+        """Extract the 'next' URL from a Canvas Link header.
+
+        Only returns URLs whose host matches the configured Canvas base URL
+        to prevent leaking the bearer token to arbitrary hosts.
+        """
+        expected_host = self._expected_host
+        for match in _NEXT_LINK_PATTERN.finditer(link_header):
+            url = match.group(1)
+            parsed_url = urlparse(url)
+            if parsed_url.hostname != expected_host:
+                raise OnyxError(
+                    OnyxErrorCode.BAD_GATEWAY,
+                    detail=(
+                        "Canvas pagination returned an unexpected host "
+                        f"({parsed_url.hostname}); expected {expected_host}"
+                    ),
+                )
+            if parsed_url.scheme != "https":
+                raise OnyxError(
+                    OnyxErrorCode.BAD_GATEWAY,
+                    detail=(
+                        "Canvas pagination link must use https, "
+                        f"got {parsed_url.scheme!r}"
+                    ),
+                )
+            return url
+        return None
+
+    def _build_headers(self) -> dict[str, str]:
+        """Return the Authorization header with the bearer token."""
+        return {"Authorization": f"Bearer {self._bearer_token}"}
+
+    def _build_url(self, endpoint: str) -> str:
+        """Build a full Canvas API URL from an endpoint path.
+
+        Assumes endpoint is non-empty (e.g. ``"courses"``, ``"announcements"``).
+        Only called on a first request, endpoint must be set for first request.
+        Verify endpoint exists in case of future changes where endpoint might be optional.
+        Leading slashes are stripped to avoid double-slash in the result.
+        self.base_url is already normalized with no trailing slash.
+        """
+        final_url = self.base_url
+        clean_endpoint = endpoint.lstrip("/")
+        if clean_endpoint:
+            final_url += "/" + clean_endpoint
+        return final_url

backend/onyx/connectors/canvas/connector.py

@@ -0,0 +1,74 @@
+from typing import Literal
+from typing import TypeAlias
+
+from pydantic import BaseModel
+
+from onyx.connectors.models import ConnectorCheckpoint
+
+
+class CanvasCourse(BaseModel):
+    id: int
+    name: str
+    course_code: str
+    created_at: str
+    workflow_state: str
+
+
+class CanvasPage(BaseModel):
+    page_id: int
+    url: str
+    title: str
+    body: str | None = None
+    created_at: str
+    updated_at: str
+    course_id: int
+
+
+class CanvasAssignment(BaseModel):
+    id: int
+    name: str
+    description: str | None = None
+    html_url: str
+    course_id: int
+    created_at: str
+    updated_at: str
+    due_at: str | None = None
+
+
+class CanvasAnnouncement(BaseModel):
+    id: int
+    title: str
+    message: str | None = None
+    html_url: str
+    posted_at: str | None = None
+    course_id: int
+
+
+CanvasStage: TypeAlias = Literal["pages", "assignments", "announcements"]
+
+
+class CanvasConnectorCheckpoint(ConnectorCheckpoint):
+    """Checkpoint state for resumable Canvas indexing.
+
+    Fields:
+        course_ids: Materialized list of course IDs to process.
+        current_course_index: Index into course_ids for current course.
+        stage: Which item type we're processing for the current course.
+        next_url: Pagination cursor within the current stage. None means
+            start from the first page; a URL means resume from that page.
+
+    Invariant:
+        If current_course_index is incremented, stage must be reset to
+        "pages" and next_url must be reset to None.
+    """
+
+    course_ids: list[int] = []
+    current_course_index: int = 0
+    stage: CanvasStage = "pages"
+    next_url: str | None = None
+
+    def advance_course(self) -> None:
+        """Move to the next course and reset within-course state."""
+        self.current_course_index += 1
+        self.stage = "pages"
+        self.next_url = None

backend/onyx/connectors/confluence/connector.py

@@ -890,8 +890,8 @@ class ConfluenceConnector(
 
     def _retrieve_all_slim_docs(
         self,
-        start: SecondsSinceUnixEpoch | None = None,  # noqa: ARG002
-        end: SecondsSinceUnixEpoch | None = None,  # noqa: ARG002
+        start: SecondsSinceUnixEpoch | None = None,
+        end: SecondsSinceUnixEpoch | None = None,
         callback: IndexingHeartbeatInterface | None = None,
         include_permissions: bool = True,
     ) -> GenerateSlimDocumentOutput:
@@ -915,8 +915,8 @@ class ConfluenceConnector(
                 self.confluence_client, doc_id, restrictions, ancestors
             ) or space_level_access_info.get(page_space_key)
 
-        # Query pages
-        page_query = self.base_cql_page_query + self.cql_label_filter
+        # Query pages (with optional time filtering for indexing_start)
+        page_query = self._construct_page_cql_query(start, end)
         for page in self.confluence_client.cql_paginate_all_expansions(
             cql=page_query,
             expand=restrictions_expand,
@@ -950,7 +950,9 @@ class ConfluenceConnector(
 
             # Query attachments for each page
             page_hierarchy_node_yielded = False
-            attachment_query = self._construct_attachment_query(_get_page_id(page))
+            attachment_query = self._construct_attachment_query(
+                _get_page_id(page), start, end
+            )
             for attachment in self.confluence_client.cql_paginate_all_expansions(
                 cql=attachment_query,
                 expand=restrictions_expand,

backend/onyx/connectors/jira/connector.py

@@ -10,6 +10,7 @@ from datetime import timedelta
 from datetime import timezone
 from typing import Any
 
+import requests
 from jira import JIRA
 from jira.exceptions import JIRAError
 from jira.resources import Issue
@@ -239,29 +240,53 @@ def enhanced_search_ids(
     )
 
 
-def bulk_fetch_issues(
-    jira_client: JIRA, issue_ids: list[str], fields: str | None = None
-) -> list[Issue]:
-    # TODO: move away from this jira library if they continue to not support
-    # the endpoints we need. Using private fields is not ideal, but
-    # is likely fine for now since we pin the library version
+def _bulk_fetch_request(
+    jira_client: JIRA, issue_ids: list[str], fields: str | None
+) -> list[dict[str, Any]]:
+    """Raw POST to the bulkfetch endpoint. Returns the list of raw issue dicts."""
     bulk_fetch_path = jira_client._get_url("issue/bulkfetch")
-
     # Prepare the payload according to Jira API v3 specification
     payload: dict[str, Any] = {"issueIdsOrKeys": issue_ids}
-
     # Only restrict fields if specified, might want to explicitly do this in the future
     # to avoid reading unnecessary data
     payload["fields"] = fields.split(",") if fields else ["*all"]
 
+    resp = jira_client._session.post(bulk_fetch_path, json=payload)
+    return resp.json()["issues"]
+
+
+def bulk_fetch_issues(
+    jira_client: JIRA, issue_ids: list[str], fields: str | None = None
+) -> list[Issue]:
+    # TODO(evan): move away from this jira library if they continue to not support
+    # the endpoints we need. Using private fields is not ideal, but
+    # is likely fine for now since we pin the library version
+
     try:
-        response = jira_client._session.post(bulk_fetch_path, json=payload).json()
+        raw_issues = _bulk_fetch_request(jira_client, issue_ids, fields)
+    except requests.exceptions.JSONDecodeError:
+        if len(issue_ids) <= 1:
+            logger.exception(
+                f"Jira bulk-fetch response for issue(s) {issue_ids} could not "
+                f"be decoded as JSON (response too large or truncated)."
+            )
+            raise
+
+        mid = len(issue_ids) // 2
+        logger.warning(
+            f"Jira bulk-fetch JSON decode failed for batch of {len(issue_ids)} issues. "
+            f"Splitting into sub-batches of {mid} and {len(issue_ids) - mid}."
+        )
+        left = bulk_fetch_issues(jira_client, issue_ids[:mid], fields)
+        right = bulk_fetch_issues(jira_client, issue_ids[mid:], fields)
+        return left + right
     except Exception as e:
         logger.error(f"Error fetching issues: {e}")
-        raise e
+        raise
+
     return [
         Issue(jira_client._options, jira_client._session, raw=issue)
-        for issue in response["issues"]
+        for issue in raw_issues
     ]
 
 

backend/onyx/connectors/sharepoint/connector.py

@@ -1765,7 +1765,11 @@ class SharepointConnector(
         checkpoint.current_drive_delta_next_link = None
         checkpoint.seen_document_ids.clear()
 
-    def _fetch_slim_documents_from_sharepoint(self) -> GenerateSlimDocumentOutput:
+    def _fetch_slim_documents_from_sharepoint(
+        self,
+        start: datetime | None = None,
+        end: datetime | None = None,
+    ) -> GenerateSlimDocumentOutput:
         site_descriptors = self._filter_excluded_sites(
             self.site_descriptors or self.fetch_sites()
         )
@@ -1786,7 +1790,9 @@ class SharepointConnector(
             # Process site documents if flag is True
             if self.include_site_documents:
                 for driveitem, drive_name, drive_web_url in self._fetch_driveitems(
-                    site_descriptor=site_descriptor
+                    site_descriptor=site_descriptor,
+                    start=start,
+                    end=end,
                 ):
                     if self._is_driveitem_excluded(driveitem):
                         logger.debug(f"Excluding by path denylist: {driveitem.web_url}")
@@ -1841,7 +1847,9 @@ class SharepointConnector(
 
             # Process site pages if flag is True
             if self.include_site_pages:
-                site_pages = self._fetch_site_pages(site_descriptor)
+                site_pages = self._fetch_site_pages(
+                    site_descriptor, start=start, end=end
+                )
                 for site_page in site_pages:
                     logger.debug(
                         f"Processing site page: {site_page.get('webUrl', site_page.get('name', 'Unknown'))}"
@@ -2565,12 +2573,22 @@ class SharepointConnector(
 
     def retrieve_all_slim_docs_perm_sync(
         self,
-        start: SecondsSinceUnixEpoch | None = None,  # noqa: ARG002
-        end: SecondsSinceUnixEpoch | None = None,  # noqa: ARG002
+        start: SecondsSinceUnixEpoch | None = None,
+        end: SecondsSinceUnixEpoch | None = None,
         callback: IndexingHeartbeatInterface | None = None,  # noqa: ARG002
     ) -> GenerateSlimDocumentOutput:
-
-        yield from self._fetch_slim_documents_from_sharepoint()
+        start_dt = (
+            datetime.fromtimestamp(start, tz=timezone.utc)
+            if start is not None
+            else None
+        )
+        end_dt = (
+            datetime.fromtimestamp(end, tz=timezone.utc) if end is not None else None
+        )
+        yield from self._fetch_slim_documents_from_sharepoint(
+            start=start_dt,
+            end=end_dt,
+        )
 
 
 if __name__ == "__main__":

backend/onyx/connectors/slack/connector.py

@@ -516,6 +516,8 @@ def _get_all_doc_ids(
     ] = default_msg_filter,
     callback: IndexingHeartbeatInterface | None = None,
     workspace_url: str | None = None,
+    start: SecondsSinceUnixEpoch | None = None,
+    end: SecondsSinceUnixEpoch | None = None,
 ) -> GenerateSlimDocumentOutput:
     """
     Get all document ids in the workspace, channel by channel
@@ -546,6 +548,8 @@ def _get_all_doc_ids(
             client=client,
             channel=channel,
             callback=callback,
+            oldest=str(start) if start else None,  # 0.0 -> None intentionally
+            latest=str(end) if end is not None else None,
         )
 
         for message_batch in channel_message_batches:
@@ -847,8 +851,8 @@ class SlackConnector(
 
     def retrieve_all_slim_docs_perm_sync(
         self,
-        start: SecondsSinceUnixEpoch | None = None,  # noqa: ARG002
-        end: SecondsSinceUnixEpoch | None = None,  # noqa: ARG002
+        start: SecondsSinceUnixEpoch | None = None,
+        end: SecondsSinceUnixEpoch | None = None,
         callback: IndexingHeartbeatInterface | None = None,
     ) -> GenerateSlimDocumentOutput:
         if self.client is None:
@@ -861,6 +865,8 @@ class SlackConnector(
             msg_filter_func=self.msg_filter_func,
             callback=callback,
             workspace_url=self._workspace_url,
+            start=start,
+            end=end,
         )
 
     def _load_from_checkpoint(

backend/onyx/db/chat.py

@@ -617,80 +617,6 @@ def reserve_message_id(
     return empty_message
 
 
-def reserve_multi_model_message_ids(
-    db_session: Session,
-    chat_session_id: UUID,
-    parent_message_id: int,
-    model_display_names: list[str],
-) -> list[ChatMessage]:
-    """Reserve N assistant message placeholders for multi-model parallel streaming.
-
-    All messages share the same parent (the user message). The parent's
-    latest_child_message_id points to the LAST reserved message so that the
-    default history-chain walker picks it up.
-    """
-    reserved: list[ChatMessage] = []
-    for display_name in model_display_names:
-        msg = ChatMessage(
-            chat_session_id=chat_session_id,
-            parent_message_id=parent_message_id,
-            latest_child_message_id=None,
-            message="Response was terminated prior to completion, try regenerating.",
-            token_count=15,  # placeholder; updated on completion by llm_loop_completion_handle
-            message_type=MessageType.ASSISTANT,
-            model_display_name=display_name,
-        )
-        db_session.add(msg)
-        reserved.append(msg)
-
-    # Flush to assign IDs without committing yet
-    db_session.flush()
-
-    # Point parent's latest_child to the last reserved message
-    parent = (
-        db_session.query(ChatMessage)
-        .filter(ChatMessage.id == parent_message_id)
-        .first()
-    )
-    if parent:
-        parent.latest_child_message_id = reserved[-1].id
-
-    db_session.commit()
-    return reserved
-
-
-def set_preferred_response(
-    db_session: Session,
-    user_message_id: int,
-    preferred_assistant_message_id: int,
-) -> None:
-    """Set the preferred assistant response for a multi-model user message.
-
-    Validates that the user message is a USER type and that the preferred
-    assistant message is a direct child of that user message.
-    """
-    user_msg = db_session.get(ChatMessage, user_message_id)
-    if user_msg is None:
-        raise ValueError(f"User message {user_message_id} not found")
-    if user_msg.message_type != MessageType.USER:
-        raise ValueError(f"Message {user_message_id} is not a user message")
-
-    assistant_msg = db_session.get(ChatMessage, preferred_assistant_message_id)
-    if assistant_msg is None:
-        raise ValueError(
-            f"Assistant message {preferred_assistant_message_id} not found"
-        )
-    if assistant_msg.parent_message_id != user_message_id:
-        raise ValueError(
-            f"Assistant message {preferred_assistant_message_id} is not a child "
-            f"of user message {user_message_id}"
-        )
-
-    user_msg.preferred_response_id = preferred_assistant_message_id
-    user_msg.latest_child_message_id = preferred_assistant_message_id
-    db_session.commit()
-
-
 def create_new_chat_message(
     chat_session_id: UUID,
     parent_message: ChatMessage,
@@ -913,8 +839,6 @@ def translate_db_message_to_chat_message_detail(
         error=chat_message.error,
         current_feedback=current_feedback,
         processing_duration_seconds=chat_message.processing_duration_seconds,
-        preferred_response_id=chat_message.preferred_response_id,
-        model_display_name=chat_message.model_display_name,
     )
 
     return chat_msg_detail

backend/onyx/db/models.py

@@ -3135,8 +3135,6 @@ class VoiceProvider(Base):
     is_default_stt: Mapped[bool] = mapped_column(Boolean, nullable=False, default=False)
     is_default_tts: Mapped[bool] = mapped_column(Boolean, nullable=False, default=False)
 
-    deleted: Mapped[bool] = mapped_column(Boolean, default=False)
-
     time_created: Mapped[datetime.datetime] = mapped_column(
         DateTime(timezone=True), server_default=func.now()
     )

backend/onyx/db/voice.py

@@ -17,39 +17,30 @@ MAX_VOICE_PLAYBACK_SPEED = 2.0
 def fetch_voice_providers(db_session: Session) -> list[VoiceProvider]:
     """Fetch all voice providers."""
     return list(
-        db_session.scalars(
-            select(VoiceProvider)
-            .where(VoiceProvider.deleted.is_(False))
-            .order_by(VoiceProvider.name)
-        ).all()
+        db_session.scalars(select(VoiceProvider).order_by(VoiceProvider.name)).all()
     )
 
 
 def fetch_voice_provider_by_id(
-    db_session: Session, provider_id: int, include_deleted: bool = False
+    db_session: Session, provider_id: int
 ) -> VoiceProvider | None:
     """Fetch a voice provider by ID."""
-    stmt = select(VoiceProvider).where(VoiceProvider.id == provider_id)
-    if not include_deleted:
-        stmt = stmt.where(VoiceProvider.deleted.is_(False))
-    return db_session.scalar(stmt)
+    return db_session.scalar(
+        select(VoiceProvider).where(VoiceProvider.id == provider_id)
+    )
 
 
 def fetch_default_stt_provider(db_session: Session) -> VoiceProvider | None:
     """Fetch the default STT provider."""
     return db_session.scalar(
-        select(VoiceProvider)
-        .where(VoiceProvider.is_default_stt.is_(True))
-        .where(VoiceProvider.deleted.is_(False))
+        select(VoiceProvider).where(VoiceProvider.is_default_stt.is_(True))
     )
 
 
 def fetch_default_tts_provider(db_session: Session) -> VoiceProvider | None:
     """Fetch the default TTS provider."""
     return db_session.scalar(
-        select(VoiceProvider)
-        .where(VoiceProvider.is_default_tts.is_(True))
-        .where(VoiceProvider.deleted.is_(False))
+        select(VoiceProvider).where(VoiceProvider.is_default_tts.is_(True))
     )
 
 
@@ -58,9 +49,7 @@ def fetch_voice_provider_by_type(
 ) -> VoiceProvider | None:
     """Fetch a voice provider by type."""
     return db_session.scalar(
-        select(VoiceProvider)
-        .where(VoiceProvider.provider_type == provider_type)
-        .where(VoiceProvider.deleted.is_(False))
+        select(VoiceProvider).where(VoiceProvider.provider_type == provider_type)
     )
 
 
@@ -119,10 +108,10 @@ def upsert_voice_provider(
 
 
 def delete_voice_provider(db_session: Session, provider_id: int) -> None:
-    """Soft-delete a voice provider by ID."""
+    """Delete a voice provider by ID."""
     provider = fetch_voice_provider_by_id(db_session, provider_id)
     if provider:
-        provider.deleted = True
+        db_session.delete(provider)
         db_session.flush()
 
 

backend/onyx/document_index/disabled.py

@@ -5,6 +5,7 @@ accidentally reaches the vector DB layer will fail loudly instead of timing
 out against a nonexistent Vespa/OpenSearch instance.
 """
 
+from collections.abc import Iterable
 from typing import Any
 
 from onyx.context.search.models import IndexFilters
@@ -66,7 +67,7 @@ class DisabledDocumentIndex(DocumentIndex):
     # ------------------------------------------------------------------
     def index(
         self,
-        chunks: list[DocMetadataAwareIndexChunk],  # noqa: ARG002
+        chunks: Iterable[DocMetadataAwareIndexChunk],  # noqa: ARG002
         index_batch_params: IndexBatchParams,  # noqa: ARG002
     ) -> set[DocumentInsertionRecord]:
         raise RuntimeError(VECTOR_DB_DISABLED_ERROR)

backend/onyx/document_index/interfaces.py

@@ -1,4 +1,5 @@
 import abc
+from collections.abc import Iterable
 from dataclasses import dataclass
 from datetime import datetime
 from typing import Any
@@ -206,7 +207,7 @@ class Indexable(abc.ABC):
     @abc.abstractmethod
     def index(
         self,
-        chunks: list[DocMetadataAwareIndexChunk],
+        chunks: Iterable[DocMetadataAwareIndexChunk],
         index_batch_params: IndexBatchParams,
     ) -> set[DocumentInsertionRecord]:
         """
@@ -226,8 +227,8 @@ class Indexable(abc.ABC):
         it is done automatically outside of this code.
 
         Parameters:
-        - chunks: Document chunks with all of the information needed for indexing to the document
-                index.
+        - chunks: Document chunks with all of the information needed for
+                indexing to the document index.
         - tenant_id: The tenant id of the user whose chunks are being indexed
         - large_chunks_enabled: Whether large chunks are enabled
 

backend/onyx/document_index/interfaces_new.py

@@ -1,4 +1,5 @@
 import abc
+from collections.abc import Iterable
 from typing import Self
 
 from pydantic import BaseModel
@@ -209,10 +210,10 @@ class Indexable(abc.ABC):
     @abc.abstractmethod
     def index(
         self,
-        chunks: list[DocMetadataAwareIndexChunk],
+        chunks: Iterable[DocMetadataAwareIndexChunk],
         indexing_metadata: IndexingMetadata,
     ) -> list[DocumentInsertionRecord]:
-        """Indexes a list of document chunks into the document index.
+        """Indexes an iterable of document chunks into the document index.
 
         This is often a batch operation including chunks from multiple
         documents.

backend/onyx/document_index/opensearch/opensearch_document_index.py

@@ -1,11 +1,12 @@
 import json
-from collections import defaultdict
+from collections.abc import Iterable
 from typing import Any
 
 import httpx
 from opensearchpy import NotFoundError
 
 from onyx.access.models import DocumentAccess
+from onyx.configs.app_configs import MAX_CHUNKS_PER_DOC_BATCH
 from onyx.configs.app_configs import VERIFY_CREATE_OPENSEARCH_INDEX_ON_INIT_MT
 from onyx.configs.chat_configs import NUM_RETURNED_HITS
 from onyx.configs.chat_configs import TITLE_CONTENT_RATIO
@@ -351,7 +352,7 @@ class OpenSearchOldDocumentIndex(OldDocumentIndex):
 
     def index(
         self,
-        chunks: list[DocMetadataAwareIndexChunk],
+        chunks: Iterable[DocMetadataAwareIndexChunk],
         index_batch_params: IndexBatchParams,
     ) -> set[OldDocumentInsertionRecord]:
         """
@@ -647,10 +648,10 @@ class OpenSearchDocumentIndex(DocumentIndex):
 
     def index(
         self,
-        chunks: list[DocMetadataAwareIndexChunk],
-        indexing_metadata: IndexingMetadata,  # noqa: ARG002
+        chunks: Iterable[DocMetadataAwareIndexChunk],
+        indexing_metadata: IndexingMetadata,
     ) -> list[DocumentInsertionRecord]:
-        """Indexes a list of document chunks into the document index.
+        """Indexes an iterable of document chunks into the document index.
 
         Groups chunks by document ID and for each document, deletes existing
         chunks and indexes the new chunks in bulk.
@@ -673,29 +674,34 @@ class OpenSearchDocumentIndex(DocumentIndex):
                 document is newly indexed or had already existed and was just
                 updated.
         """
-        # Group chunks by document ID.
-        doc_id_to_chunks: dict[str, list[DocMetadataAwareIndexChunk]] = defaultdict(
-            list
+        total_chunks = sum(
+            cc.new_chunk_cnt
+            for cc in indexing_metadata.doc_id_to_chunk_cnt_diff.values()
         )
-        for chunk in chunks:
-            doc_id_to_chunks[chunk.source_document.id].append(chunk)
         logger.debug(
-            f"[OpenSearchDocumentIndex] Indexing {len(chunks)} chunks from {len(doc_id_to_chunks)} "
+            f"[OpenSearchDocumentIndex] Indexing {total_chunks} chunks from {len(indexing_metadata.doc_id_to_chunk_cnt_diff)} "
             f"documents for index {self._index_name}."
         )
 
         document_indexing_results: list[DocumentInsertionRecord] = []
-        # Try to index per-document.
-        for _, chunks in doc_id_to_chunks.items():
+        deleted_doc_ids: set[str] = set()
+        # Buffer chunks per document as they arrive from the iterable.
+        # When the document ID changes flush the buffered chunks.
+        current_doc_id: str | None = None
+        current_chunks: list[DocMetadataAwareIndexChunk] = []
+
+        def _flush_chunks(doc_chunks: list[DocMetadataAwareIndexChunk]) -> None:
+            assert len(doc_chunks) > 0, "doc_chunks is empty"
+
             # Create a batch of OpenSearch-formatted chunks for bulk insertion.
-            # Do this before deleting existing chunks to reduce the amount of
-            # time the document index has no content for a given document, and
-            # to reduce the chance of entering a state where we delete chunks,
-            # then some error happens, and never successfully index new chunks.
+            # Since we are doing this in batches, an error occurring midway
+            # can result in a state where chunks are deleted and not all the
+            # new chunks have been indexed.
             chunk_batch: list[DocumentChunk] = [
-                _convert_onyx_chunk_to_opensearch_document(chunk) for chunk in chunks
+                _convert_onyx_chunk_to_opensearch_document(chunk)
+                for chunk in doc_chunks
             ]
-            onyx_document: Document = chunks[0].source_document
+            onyx_document: Document = doc_chunks[0].source_document
             # First delete the doc's chunks from the index. This is so that
             # there are no dangling chunks in the index, in the event that the
             # new document's content contains fewer chunks than the previous
@@ -704,22 +710,43 @@ class OpenSearchDocumentIndex(DocumentIndex):
             # if the chunk count has actually decreased. This assumes that
             # overlapping chunks are perfectly overwritten. If we can't
             # guarantee that then we need the code as-is.
-            num_chunks_deleted = self.delete(
-                onyx_document.id, onyx_document.chunk_count
-            )
-            # If we see that chunks were deleted we assume the doc already
-            # existed.
-            document_insertion_record = DocumentInsertionRecord(
-                document_id=onyx_document.id,
-                already_existed=num_chunks_deleted > 0,
-            )
+            if onyx_document.id not in deleted_doc_ids:
+                num_chunks_deleted = self.delete(
+                    onyx_document.id, onyx_document.chunk_count
+                )
+                deleted_doc_ids.add(onyx_document.id)
+                # If we see that chunks were deleted we assume the doc already
+                # existed. We record the result before bulk_index_documents
+                # runs. If indexing raises, this entire result list is discarded
+                # by the caller's retry logic, so early recording is safe.
+                document_indexing_results.append(
+                    DocumentInsertionRecord(
+                        document_id=onyx_document.id,
+                        already_existed=num_chunks_deleted > 0,
+                    )
+                )
             # Now index. This will raise if a chunk of the same ID exists, which
             # we do not expect because we should have deleted all chunks.
             self._client.bulk_index_documents(
                 documents=chunk_batch,
                 tenant_state=self._tenant_state,
             )
-            document_indexing_results.append(document_insertion_record)
+
+        for chunk in chunks:
+            doc_id = chunk.source_document.id
+            if doc_id != current_doc_id:
+                if current_chunks:
+                    _flush_chunks(current_chunks)
+                current_doc_id = doc_id
+                current_chunks = [chunk]
+            elif len(current_chunks) >= MAX_CHUNKS_PER_DOC_BATCH:
+                _flush_chunks(current_chunks)
+                current_chunks = [chunk]
+            else:
+                current_chunks.append(chunk)
+
+        if current_chunks:
+            _flush_chunks(current_chunks)
 
         return document_indexing_results
 

backend/onyx/document_index/vespa/index.py

@@ -6,6 +6,7 @@ import re
 import time
 import urllib
 import zipfile
+from collections.abc import Iterable
 from dataclasses import dataclass
 from datetime import datetime
 from datetime import timedelta
@@ -461,7 +462,7 @@ class VespaIndex(DocumentIndex):
 
     def index(
         self,
-        chunks: list[DocMetadataAwareIndexChunk],
+        chunks: Iterable[DocMetadataAwareIndexChunk],
         index_batch_params: IndexBatchParams,
     ) -> set[OldDocumentInsertionRecord]:
         """

backend/onyx/document_index/vespa/vespa_document_index.py

@@ -1,6 +1,8 @@
 import concurrent.futures
 import logging
 import random
+from collections.abc import Generator
+from collections.abc import Iterable
 from typing import Any
 from uuid import UUID
 
@@ -8,6 +10,7 @@ import httpx
 from pydantic import BaseModel
 from retry import retry
 
+from onyx.configs.app_configs import MAX_CHUNKS_PER_DOC_BATCH
 from onyx.configs.app_configs import RECENCY_BIAS_MULTIPLIER
 from onyx.configs.app_configs import RERANK_COUNT
 from onyx.configs.chat_configs import DOC_TIME_DECAY
@@ -318,7 +321,7 @@ class VespaDocumentIndex(DocumentIndex):
 
     def index(
         self,
-        chunks: list[DocMetadataAwareIndexChunk],
+        chunks: Iterable[DocMetadataAwareIndexChunk],
         indexing_metadata: IndexingMetadata,
     ) -> list[DocumentInsertionRecord]:
         doc_id_to_chunk_cnt_diff = indexing_metadata.doc_id_to_chunk_cnt_diff
@@ -338,22 +341,31 @@ class VespaDocumentIndex(DocumentIndex):
 
         # Vespa has restrictions on valid characters, yet document IDs come from
         # external w.r.t. this class. We need to sanitize them.
-        cleaned_chunks: list[DocMetadataAwareIndexChunk] = [
-            clean_chunk_id_copy(chunk) for chunk in chunks
-        ]
-        assert len(cleaned_chunks) == len(
-            chunks
-        ), "Bug: Cleaned chunks and input chunks have different lengths."
+        #
+        # Instead of materializing all cleaned chunks upfront, we stream them
+        # through a generator that cleans IDs and builds the original-ID mapping
+        # incrementally as chunks flow into Vespa.
+        def _clean_and_track(
+            chunks_iter: Iterable[DocMetadataAwareIndexChunk],
+            id_map: dict[str, str],
+            seen_ids: set[str],
+        ) -> Generator[DocMetadataAwareIndexChunk, None, None]:
+            """Cleans chunk IDs and builds the original-ID mapping
+            incrementally as chunks flow through, avoiding a separate
+            materialization pass."""
+            for chunk in chunks_iter:
+                original_id = chunk.source_document.id
+                cleaned = clean_chunk_id_copy(chunk)
+                cleaned_id = cleaned.source_document.id
+                # Needed so the final DocumentInsertionRecord returned can have
+                # the original document ID. cleaned_chunks might not contain IDs
+                # exactly as callers supplied them.
+                id_map[cleaned_id] = original_id
+                seen_ids.add(cleaned_id)
+                yield cleaned
 
-        # Needed so the final DocumentInsertionRecord returned can have the
-        # original document ID. cleaned_chunks might not contain IDs exactly as
-        # callers supplied them.
-        new_document_id_to_original_document_id: dict[str, str] = dict()
-        for i, cleaned_chunk in enumerate(cleaned_chunks):
-            old_chunk = chunks[i]
-            new_document_id_to_original_document_id[
-                cleaned_chunk.source_document.id
-            ] = old_chunk.source_document.id
+        new_document_id_to_original_document_id: dict[str, str] = {}
+        all_cleaned_doc_ids: set[str] = set()
 
         existing_docs: set[str] = set()
 
@@ -409,8 +421,16 @@ class VespaDocumentIndex(DocumentIndex):
                     executor=executor,
                 )
 
-            # Insert new Vespa documents.
-            for chunk_batch in batch_generator(cleaned_chunks, BATCH_SIZE):
+            # Insert new Vespa documents, streaming through the cleaning
+            # pipeline so chunks are never fully materialized.
+            cleaned_chunks = _clean_and_track(
+                chunks,
+                new_document_id_to_original_document_id,
+                all_cleaned_doc_ids,
+            )
+            for chunk_batch in batch_generator(
+                cleaned_chunks, min(BATCH_SIZE, MAX_CHUNKS_PER_DOC_BATCH)
+            ):
                 batch_index_vespa_chunks(
                     chunks=chunk_batch,
                     index_name=self._index_name,
@@ -419,10 +439,6 @@ class VespaDocumentIndex(DocumentIndex):
                     executor=executor,
                 )
 
-        all_cleaned_doc_ids: set[str] = {
-            chunk.source_document.id for chunk in cleaned_chunks
-        }
-
         return [
             DocumentInsertionRecord(
                 document_id=new_document_id_to_original_document_id[cleaned_doc_id],

backend/onyx/file_processing/extract_file_text.py

@@ -44,6 +44,7 @@ KNOWN_OPENPYXL_BUGS = [
     "Value must be either numerical or a string containing a wildcard",
     "File contains no valid workbook part",
     "Unable to read workbook: could not read stylesheet from None",
+    "Colors must be aRGB hex values",
 ]
 
 

backend/onyx/indexing/adapters/document_indexing_adapter.py

@@ -19,7 +19,8 @@ from onyx.db.document import update_docs_updated_at__no_commit
 from onyx.db.document_set import fetch_document_sets_for_documents
 from onyx.indexing.indexing_pipeline import DocumentBatchPrepareContext
 from onyx.indexing.indexing_pipeline import index_doc_batch_prepare
-from onyx.indexing.models import BuildMetadataAwareChunksResult
+from onyx.indexing.models import ChunkEnrichmentContext
+from onyx.indexing.models import DocAwareChunk
 from onyx.indexing.models import DocMetadataAwareIndexChunk
 from onyx.indexing.models import IndexChunk
 from onyx.indexing.models import UpdatableChunkData
@@ -85,14 +86,21 @@ class DocumentIndexingBatchAdapter:
         ) as transaction:
             yield transaction
 
-    def build_metadata_aware_chunks(
+    def prepare_enrichment(
         self,
-        chunks_with_embeddings: list[IndexChunk],
-        chunk_content_scores: list[float],
-        tenant_id: str,
         context: DocumentBatchPrepareContext,
-    ) -> BuildMetadataAwareChunksResult:
-        """Enrich chunks with access, document sets, boosts, token counts, and hierarchy."""
+        tenant_id: str,
+        chunks: list[DocAwareChunk],
+    ) -> "DocumentChunkEnricher":
+        """Do all DB lookups once and return a per-chunk enricher."""
+        updatable_ids = [doc.id for doc in context.updatable_docs]
+
+        doc_id_to_new_chunk_cnt: dict[str, int] = {
+            doc_id: 0 for doc_id in updatable_ids
+        }
+        for chunk in chunks:
+            if chunk.source_document.id in doc_id_to_new_chunk_cnt:
+                doc_id_to_new_chunk_cnt[chunk.source_document.id] += 1
 
         no_access = DocumentAccess.build(
             user_emails=[],
@@ -102,67 +110,30 @@ class DocumentIndexingBatchAdapter:
             is_public=False,
         )
 
-        updatable_ids = [doc.id for doc in context.updatable_docs]
-
-        doc_id_to_access_info = get_access_for_documents(
-            document_ids=updatable_ids, db_session=self.db_session
-        )
-        doc_id_to_document_set = {
-            document_id: document_sets
-            for document_id, document_sets in fetch_document_sets_for_documents(
+        return DocumentChunkEnricher(
+            doc_id_to_access_info=get_access_for_documents(
                 document_ids=updatable_ids, db_session=self.db_session
-            )
-        }
-
-        doc_id_to_previous_chunk_cnt: dict[str, int] = {
-            document_id: chunk_count
-            for document_id, chunk_count in fetch_chunk_counts_for_documents(
-                document_ids=updatable_ids,
-                db_session=self.db_session,
-            )
-        }
-
-        doc_id_to_new_chunk_cnt: dict[str, int] = {
-            doc_id: 0 for doc_id in updatable_ids
-        }
-        for chunk in chunks_with_embeddings:
-            if chunk.source_document.id in doc_id_to_new_chunk_cnt:
-                doc_id_to_new_chunk_cnt[chunk.source_document.id] += 1
-
-        # Get ancestor hierarchy node IDs for each document
-        doc_id_to_ancestor_ids = self._get_ancestor_ids_for_documents(
-            context.updatable_docs, tenant_id
-        )
-
-        access_aware_chunks = [
-            DocMetadataAwareIndexChunk.from_index_chunk(
-                index_chunk=chunk,
-                access=doc_id_to_access_info.get(chunk.source_document.id, no_access),
-                document_sets=set(
-                    doc_id_to_document_set.get(chunk.source_document.id, [])
-                ),
-                user_project=[],
-                personas=[],
-                boost=(
-                    context.id_to_boost_map[chunk.source_document.id]
-                    if chunk.source_document.id in context.id_to_boost_map
-                    else DEFAULT_BOOST
-                ),
-                tenant_id=tenant_id,
-                aggregated_chunk_boost_factor=chunk_content_scores[chunk_num],
-                ancestor_hierarchy_node_ids=doc_id_to_ancestor_ids[
-                    chunk.source_document.id
-                ],
-            )
-            for chunk_num, chunk in enumerate(chunks_with_embeddings)
-        ]
-
-        return BuildMetadataAwareChunksResult(
-            chunks=access_aware_chunks,
-            doc_id_to_previous_chunk_cnt=doc_id_to_previous_chunk_cnt,
-            doc_id_to_new_chunk_cnt=doc_id_to_new_chunk_cnt,
-            user_file_id_to_raw_text={},
-            user_file_id_to_token_count={},
+            ),
+            doc_id_to_document_set={
+                document_id: document_sets
+                for document_id, document_sets in fetch_document_sets_for_documents(
+                    document_ids=updatable_ids, db_session=self.db_session
+                )
+            },
+            doc_id_to_ancestor_ids=self._get_ancestor_ids_for_documents(
+                context.updatable_docs, tenant_id
+            ),
+            id_to_boost_map=context.id_to_boost_map,
+            doc_id_to_previous_chunk_cnt={
+                document_id: chunk_count
+                for document_id, chunk_count in fetch_chunk_counts_for_documents(
+                    document_ids=updatable_ids,
+                    db_session=self.db_session,
+                )
+            },
+            doc_id_to_new_chunk_cnt=dict(doc_id_to_new_chunk_cnt),
+            no_access=no_access,
+            tenant_id=tenant_id,
         )
 
     def _get_ancestor_ids_for_documents(
@@ -203,7 +174,7 @@ class DocumentIndexingBatchAdapter:
         context: DocumentBatchPrepareContext,
         updatable_chunk_data: list[UpdatableChunkData],
         filtered_documents: list[Document],
-        result: BuildMetadataAwareChunksResult,
+        enrichment: ChunkEnrichmentContext,
     ) -> None:
         """Finalize DB updates, store plaintext, and mark docs as indexed."""
         updatable_ids = [doc.id for doc in context.updatable_docs]
@@ -227,7 +198,7 @@ class DocumentIndexingBatchAdapter:
 
         update_docs_chunk_count__no_commit(
             document_ids=updatable_ids,
-            doc_id_to_chunk_count=result.doc_id_to_new_chunk_cnt,
+            doc_id_to_chunk_count=enrichment.doc_id_to_new_chunk_cnt,
             db_session=self.db_session,
         )
 
@@ -249,3 +220,52 @@ class DocumentIndexingBatchAdapter:
         )
 
         self.db_session.commit()
+
+
+class DocumentChunkEnricher:
+    """Pre-computed metadata for per-chunk enrichment of connector documents."""
+
+    def __init__(
+        self,
+        doc_id_to_access_info: dict[str, DocumentAccess],
+        doc_id_to_document_set: dict[str, list[str]],
+        doc_id_to_ancestor_ids: dict[str, list[int]],
+        id_to_boost_map: dict[str, int],
+        doc_id_to_previous_chunk_cnt: dict[str, int],
+        doc_id_to_new_chunk_cnt: dict[str, int],
+        no_access: DocumentAccess,
+        tenant_id: str,
+    ) -> None:
+        self._doc_id_to_access_info = doc_id_to_access_info
+        self._doc_id_to_document_set = doc_id_to_document_set
+        self._doc_id_to_ancestor_ids = doc_id_to_ancestor_ids
+        self._id_to_boost_map = id_to_boost_map
+        self._no_access = no_access
+        self._tenant_id = tenant_id
+        self.doc_id_to_previous_chunk_cnt = doc_id_to_previous_chunk_cnt
+        self.doc_id_to_new_chunk_cnt = doc_id_to_new_chunk_cnt
+
+    def enrich_chunk(
+        self, chunk: IndexChunk, score: float
+    ) -> DocMetadataAwareIndexChunk:
+        return DocMetadataAwareIndexChunk.from_index_chunk(
+            index_chunk=chunk,
+            access=self._doc_id_to_access_info.get(
+                chunk.source_document.id, self._no_access
+            ),
+            document_sets=set(
+                self._doc_id_to_document_set.get(chunk.source_document.id, [])
+            ),
+            user_project=[],
+            personas=[],
+            boost=(
+                self._id_to_boost_map[chunk.source_document.id]
+                if chunk.source_document.id in self._id_to_boost_map
+                else DEFAULT_BOOST
+            ),
+            tenant_id=self._tenant_id,
+            aggregated_chunk_boost_factor=score,
+            ancestor_hierarchy_node_ids=self._doc_id_to_ancestor_ids[
+                chunk.source_document.id
+            ],
+        )

backend/onyx/indexing/adapters/user_file_indexing_adapter.py

@@ -1,6 +1,9 @@
+from __future__ import annotations
+
 import contextlib
 import datetime
 import time
+from collections import defaultdict
 from collections.abc import Generator
 from uuid import UUID
 
@@ -24,11 +27,13 @@ from onyx.db.user_file import fetch_persona_ids_for_user_files
 from onyx.db.user_file import fetch_user_project_ids_for_user_files
 from onyx.file_store.utils import store_user_file_plaintext
 from onyx.indexing.indexing_pipeline import DocumentBatchPrepareContext
-from onyx.indexing.models import BuildMetadataAwareChunksResult
+from onyx.indexing.models import ChunkEnrichmentContext
+from onyx.indexing.models import DocAwareChunk
 from onyx.indexing.models import DocMetadataAwareIndexChunk
 from onyx.indexing.models import IndexChunk
 from onyx.indexing.models import UpdatableChunkData
 from onyx.llm.factory import get_default_llm
+from onyx.natural_language_processing.utils import count_tokens
 from onyx.natural_language_processing.utils import get_tokenizer
 from onyx.utils.logger import setup_logger
 
@@ -101,13 +106,20 @@ class UserFileIndexingAdapter:
                 f"Failed to acquire locks after {_NUM_LOCK_ATTEMPTS} attempts for user files: {[doc.id for doc in documents]}"
             )
 
-    def build_metadata_aware_chunks(
+    def prepare_enrichment(
         self,
-        chunks_with_embeddings: list[IndexChunk],
-        chunk_content_scores: list[float],
-        tenant_id: str,
         context: DocumentBatchPrepareContext,
-    ) -> BuildMetadataAwareChunksResult:
+        tenant_id: str,
+        chunks: list[DocAwareChunk],
+    ) -> UserFileChunkEnricher:
+        """Do all DB lookups and pre-compute file metadata from chunks."""
+        updatable_ids = [doc.id for doc in context.updatable_docs]
+
+        doc_id_to_new_chunk_cnt: dict[str, int] = defaultdict(int)
+        content_by_file: dict[str, list[str]] = defaultdict(list)
+        for chunk in chunks:
+            doc_id_to_new_chunk_cnt[chunk.source_document.id] += 1
+            content_by_file[chunk.source_document.id].append(chunk.content)
 
         no_access = DocumentAccess.build(
             user_emails=[],
@@ -117,7 +129,6 @@ class UserFileIndexingAdapter:
             is_public=False,
         )
 
-        updatable_ids = [doc.id for doc in context.updatable_docs]
         user_file_id_to_project_ids = fetch_user_project_ids_for_user_files(
             user_file_ids=updatable_ids,
             db_session=self.db_session,
@@ -138,17 +149,6 @@ class UserFileIndexingAdapter:
             )
         }
 
-        user_file_id_to_new_chunk_cnt: dict[str, int] = {
-            user_file_id: len(
-                [
-                    chunk
-                    for chunk in chunks_with_embeddings
-                    if chunk.source_document.id == user_file_id
-                ]
-            )
-            for user_file_id in updatable_ids
-        }
-
         # Initialize tokenizer used for token count calculation
         try:
             llm = get_default_llm()
@@ -163,46 +163,30 @@ class UserFileIndexingAdapter:
         user_file_id_to_raw_text: dict[str, str] = {}
         user_file_id_to_token_count: dict[str, int | None] = {}
         for user_file_id in updatable_ids:
-            user_file_chunks = [
-                chunk
-                for chunk in chunks_with_embeddings
-                if chunk.source_document.id == user_file_id
-            ]
-            if user_file_chunks:
-                combined_content = " ".join(
-                    [chunk.content for chunk in user_file_chunks]
-                )
+            contents = content_by_file.get(user_file_id)
+            if contents:
+                combined_content = " ".join(contents)
                 user_file_id_to_raw_text[str(user_file_id)] = combined_content
-                token_count = (
-                    len(llm_tokenizer.encode(combined_content)) if llm_tokenizer else 0
+                token_count: int = (
+                    count_tokens(combined_content, llm_tokenizer)
+                    if llm_tokenizer
+                    else 0
                 )
                 user_file_id_to_token_count[str(user_file_id)] = token_count
             else:
                 user_file_id_to_raw_text[str(user_file_id)] = ""
                 user_file_id_to_token_count[str(user_file_id)] = None
 
-        access_aware_chunks = [
-            DocMetadataAwareIndexChunk.from_index_chunk(
-                index_chunk=chunk,
-                access=user_file_id_to_access.get(chunk.source_document.id, no_access),
-                document_sets=set(),
-                user_project=user_file_id_to_project_ids.get(
-                    chunk.source_document.id, []
-                ),
-                personas=user_file_id_to_persona_ids.get(chunk.source_document.id, []),
-                boost=DEFAULT_BOOST,
-                tenant_id=tenant_id,
-                aggregated_chunk_boost_factor=chunk_content_scores[chunk_num],
-            )
-            for chunk_num, chunk in enumerate(chunks_with_embeddings)
-        ]
-
-        return BuildMetadataAwareChunksResult(
-            chunks=access_aware_chunks,
+        return UserFileChunkEnricher(
+            user_file_id_to_access=user_file_id_to_access,
+            user_file_id_to_project_ids=user_file_id_to_project_ids,
+            user_file_id_to_persona_ids=user_file_id_to_persona_ids,
             doc_id_to_previous_chunk_cnt=user_file_id_to_previous_chunk_cnt,
-            doc_id_to_new_chunk_cnt=user_file_id_to_new_chunk_cnt,
+            doc_id_to_new_chunk_cnt=dict(doc_id_to_new_chunk_cnt),
             user_file_id_to_raw_text=user_file_id_to_raw_text,
             user_file_id_to_token_count=user_file_id_to_token_count,
+            no_access=no_access,
+            tenant_id=tenant_id,
         )
 
     def _notify_assistant_owners_if_files_ready(
@@ -246,8 +230,9 @@ class UserFileIndexingAdapter:
         context: DocumentBatchPrepareContext,
         updatable_chunk_data: list[UpdatableChunkData],  # noqa: ARG002
         filtered_documents: list[Document],  # noqa: ARG002
-        result: BuildMetadataAwareChunksResult,
+        enrichment: ChunkEnrichmentContext,
     ) -> None:
+        assert isinstance(enrichment, UserFileChunkEnricher)
         user_file_ids = [doc.id for doc in context.updatable_docs]
 
         user_files = (
@@ -263,8 +248,10 @@ class UserFileIndexingAdapter:
             user_file.last_project_sync_at = datetime.datetime.now(
                 datetime.timezone.utc
             )
-            user_file.chunk_count = result.doc_id_to_new_chunk_cnt[str(user_file.id)]
-            user_file.token_count = result.user_file_id_to_token_count[
+            user_file.chunk_count = enrichment.doc_id_to_new_chunk_cnt.get(
+                str(user_file.id), 0
+            )
+            user_file.token_count = enrichment.user_file_id_to_token_count[
                 str(user_file.id)
             ]
 
@@ -276,8 +263,54 @@ class UserFileIndexingAdapter:
         # Store the plaintext in the file store for faster retrieval
         # NOTE: this creates its own session to avoid committing the overall
         # transaction.
-        for user_file_id, raw_text in result.user_file_id_to_raw_text.items():
+        for user_file_id, raw_text in enrichment.user_file_id_to_raw_text.items():
             store_user_file_plaintext(
                 user_file_id=UUID(user_file_id),
                 plaintext_content=raw_text,
             )
+
+
+class UserFileChunkEnricher:
+    """Pre-computed metadata for per-chunk enrichment of user-uploaded files."""
+
+    def __init__(
+        self,
+        user_file_id_to_access: dict[str, DocumentAccess],
+        user_file_id_to_project_ids: dict[str, list[int]],
+        user_file_id_to_persona_ids: dict[str, list[int]],
+        doc_id_to_previous_chunk_cnt: dict[str, int],
+        doc_id_to_new_chunk_cnt: dict[str, int],
+        user_file_id_to_raw_text: dict[str, str],
+        user_file_id_to_token_count: dict[str, int | None],
+        no_access: DocumentAccess,
+        tenant_id: str,
+    ) -> None:
+        self._user_file_id_to_access = user_file_id_to_access
+        self._user_file_id_to_project_ids = user_file_id_to_project_ids
+        self._user_file_id_to_persona_ids = user_file_id_to_persona_ids
+        self._no_access = no_access
+        self._tenant_id = tenant_id
+        self.doc_id_to_previous_chunk_cnt = doc_id_to_previous_chunk_cnt
+        self.doc_id_to_new_chunk_cnt = doc_id_to_new_chunk_cnt
+        self.user_file_id_to_raw_text = user_file_id_to_raw_text
+        self.user_file_id_to_token_count = user_file_id_to_token_count
+
+    def enrich_chunk(
+        self, chunk: IndexChunk, score: float
+    ) -> DocMetadataAwareIndexChunk:
+        return DocMetadataAwareIndexChunk.from_index_chunk(
+            index_chunk=chunk,
+            access=self._user_file_id_to_access.get(
+                chunk.source_document.id, self._no_access
+            ),
+            document_sets=set(),
+            user_project=self._user_file_id_to_project_ids.get(
+                chunk.source_document.id, []
+            ),
+            personas=self._user_file_id_to_persona_ids.get(
+                chunk.source_document.id, []
+            ),
+            boost=DEFAULT_BOOST,
+            tenant_id=self._tenant_id,
+            aggregated_chunk_boost_factor=score,
+        )

backend/onyx/indexing/chunk_batch_store.py

@@ -0,0 +1,89 @@
+import pickle
+import shutil
+import tempfile
+from collections.abc import Iterator
+from pathlib import Path
+
+from onyx.indexing.models import IndexChunk
+
+
+class ChunkBatchStore:
+    """Manages serialization of embedded chunks to a temporary directory.
+
+    Owns the temp directory lifetime and provides save/load/stream/scrub
+    operations.
+
+    Use as a context manager to ensure cleanup::
+
+        with ChunkBatchStore() as store:
+            store.save(chunks, batch_idx=0)
+            for chunk in store.stream():
+                ...
+    """
+
+    _EXT = ".pkl"
+
+    def __init__(self) -> None:
+        self._tmpdir: Path | None = None
+
+    # -- context manager -----------------------------------------------------
+
+    def __enter__(self) -> "ChunkBatchStore":
+        self._tmpdir = Path(tempfile.mkdtemp(prefix="onyx_embeddings_"))
+        return self
+
+    def __exit__(self, *_exc: object) -> None:
+        if self._tmpdir is not None:
+            shutil.rmtree(self._tmpdir, ignore_errors=True)
+            self._tmpdir = None
+
+    @property
+    def _dir(self) -> Path:
+        assert self._tmpdir is not None, "ChunkBatchStore used outside context manager"
+        return self._tmpdir
+
+    # -- storage primitives --------------------------------------------------
+
+    def save(self, chunks: list[IndexChunk], batch_idx: int) -> None:
+        """Serialize a batch of embedded chunks to disk."""
+        with open(self._dir / f"batch_{batch_idx}{self._EXT}", "wb") as f:
+            pickle.dump(chunks, f)
+
+    def _load(self, batch_file: Path) -> list[IndexChunk]:
+        """Deserialize a batch of embedded chunks from a file."""
+        with open(batch_file, "rb") as f:
+            return pickle.load(f)
+
+    def _batch_files(self) -> list[Path]:
+        """Return batch files sorted by numeric index."""
+        return sorted(
+            self._dir.glob(f"batch_*{self._EXT}"),
+            key=lambda p: int(p.stem.removeprefix("batch_")),
+        )
+
+    # -- higher-level operations ---------------------------------------------
+
+    def stream(self) -> Iterator[IndexChunk]:
+        """Yield all chunks across all batch files.
+
+        Each call returns a fresh generator, so the data can be iterated
+        multiple times (e.g. once per document index).
+        """
+        for batch_file in self._batch_files():
+            yield from self._load(batch_file)
+
+    def scrub_failed_docs(self, failed_doc_ids: set[str]) -> None:
+        """Remove chunks belonging to *failed_doc_ids* from all batch files.
+
+        When a document fails embedding in batch N, earlier batches may
+        already contain successfully embedded chunks for that document.
+        This ensures the output is all-or-nothing per document.
+        """
+        for batch_file in self._batch_files():
+            batch_chunks = self._load(batch_file)
+            cleaned = [
+                c for c in batch_chunks if c.source_document.id not in failed_doc_ids
+            ]
+            if len(cleaned) != len(batch_chunks):
+                with open(batch_file, "wb") as f:
+                    pickle.dump(cleaned, f)

backend/onyx/indexing/indexing_pipeline.py

@@ -1,5 +1,8 @@
 from collections import defaultdict
 from collections.abc import Callable
+from collections.abc import Generator
+from collections.abc import Iterator
+from contextlib import contextmanager
 from typing import Protocol
 
 from pydantic import BaseModel
@@ -9,6 +12,7 @@ from sqlalchemy.orm import Session
 from onyx.configs.app_configs import DEFAULT_CONTEXTUAL_RAG_LLM_NAME
 from onyx.configs.app_configs import DEFAULT_CONTEXTUAL_RAG_LLM_PROVIDER
 from onyx.configs.app_configs import ENABLE_CONTEXTUAL_RAG
+from onyx.configs.app_configs import MAX_CHUNKS_PER_DOC_BATCH
 from onyx.configs.app_configs import MAX_DOCUMENT_CHARS
 from onyx.configs.app_configs import MAX_TOKENS_FOR_FULL_INCLUSION
 from onyx.configs.app_configs import USE_CHUNK_SUMMARY
@@ -43,10 +47,12 @@ from onyx.document_index.interfaces import DocumentMetadata
 from onyx.document_index.interfaces import IndexBatchParams
 from onyx.file_processing.image_summarization import summarize_image_with_error_handling
 from onyx.file_store.file_store import get_default_file_store
+from onyx.indexing.chunk_batch_store import ChunkBatchStore
 from onyx.indexing.chunker import Chunker
 from onyx.indexing.embedder import embed_chunks_with_failure_handling
 from onyx.indexing.embedder import IndexingEmbedder
 from onyx.indexing.models import DocAwareChunk
+from onyx.indexing.models import DocMetadataAwareIndexChunk
 from onyx.indexing.models import IndexingBatchAdapter
 from onyx.indexing.models import UpdatableChunkData
 from onyx.indexing.vector_db_insertion import write_chunks_to_vector_db_with_backoff
@@ -63,6 +69,7 @@ from onyx.natural_language_processing.utils import tokenizer_trim_middle
 from onyx.prompts.contextual_retrieval import CONTEXTUAL_RAG_PROMPT1
 from onyx.prompts.contextual_retrieval import CONTEXTUAL_RAG_PROMPT2
 from onyx.prompts.contextual_retrieval import DOCUMENT_SUMMARY_PROMPT
+from onyx.utils.batching import batch_generator
 from onyx.utils.logger import setup_logger
 from onyx.utils.postgres_sanitization import sanitize_documents_for_postgres
 from onyx.utils.threadpool_concurrency import run_functions_tuples_in_parallel
@@ -91,6 +98,20 @@ class IndexingPipelineResult(BaseModel):
 
     failures: list[ConnectorFailure]
 
+    @classmethod
+    def empty(cls, total_docs: int) -> "IndexingPipelineResult":
+        return cls(
+            new_docs=0,
+            total_docs=total_docs,
+            total_chunks=0,
+            failures=[],
+        )
+
+
+class ChunkEmbeddingResult(BaseModel):
+    successful_chunk_ids: list[tuple[int, str]]  # (chunk_id, document_id)
+    connector_failures: list[ConnectorFailure]
+
 
 class IndexingPipelineProtocol(Protocol):
     def __call__(
@@ -139,6 +160,110 @@ def _upsert_documents_in_db(
         )
 
 
+def _get_failed_doc_ids(failures: list[ConnectorFailure]) -> set[str]:
+    """Extract document IDs from a list of connector failures."""
+    return {f.failed_document.document_id for f in failures if f.failed_document}
+
+
+def _embed_chunks_to_store(
+    chunks: list[DocAwareChunk],
+    embedder: IndexingEmbedder,
+    tenant_id: str,
+    request_id: str | None,
+    store: ChunkBatchStore,
+) -> ChunkEmbeddingResult:
+    """Embed chunks in batches, spilling each batch to *store*.
+
+    If a document fails embedding in any batch, its chunks are excluded from
+    all batches (including earlier ones already written) so that the output
+    is all-or-nothing per document.
+    """
+    successful_chunk_ids: list[tuple[int, str]] = []
+    all_embedding_failures: list[ConnectorFailure] = []
+    # Track failed doc IDs across all batches so that a failure in batch N
+    # causes chunks for that doc to be skipped in batch N+1 and stripped
+    # from earlier batches.
+    all_failed_doc_ids: set[str] = set()
+
+    for batch_idx, chunk_batch in enumerate(
+        batch_generator(chunks, MAX_CHUNKS_PER_DOC_BATCH)
+    ):
+        # Skip chunks belonging to documents that failed in earlier batches.
+        chunk_batch = [
+            c for c in chunk_batch if c.source_document.id not in all_failed_doc_ids
+        ]
+        if not chunk_batch:
+            continue
+
+        logger.debug(f"Embedding batch {batch_idx}: {len(chunk_batch)} chunks")
+
+        chunks_with_embeddings, embedding_failures = embed_chunks_with_failure_handling(
+            chunks=chunk_batch,
+            embedder=embedder,
+            tenant_id=tenant_id,
+            request_id=request_id,
+        )
+        all_embedding_failures.extend(embedding_failures)
+        all_failed_doc_ids.update(_get_failed_doc_ids(embedding_failures))
+
+        # Only keep successfully embedded chunks for non-failed docs.
+        chunks_with_embeddings = [
+            c
+            for c in chunks_with_embeddings
+            if c.source_document.id not in all_failed_doc_ids
+        ]
+
+        successful_chunk_ids.extend(
+            (c.chunk_id, c.source_document.id) for c in chunks_with_embeddings
+        )
+
+        store.save(chunks_with_embeddings, batch_idx)
+        del chunks_with_embeddings
+
+    # Scrub earlier batches for docs that failed in later batches.
+    if all_failed_doc_ids:
+        store.scrub_failed_docs(all_failed_doc_ids)
+        successful_chunk_ids = [
+            (chunk_id, doc_id)
+            for chunk_id, doc_id in successful_chunk_ids
+            if doc_id not in all_failed_doc_ids
+        ]
+
+    return ChunkEmbeddingResult(
+        successful_chunk_ids=successful_chunk_ids,
+        connector_failures=all_embedding_failures,
+    )
+
+
+@contextmanager
+def embed_and_stream(
+    chunks: list[DocAwareChunk],
+    embedder: IndexingEmbedder,
+    tenant_id: str,
+    request_id: str | None,
+) -> Generator[tuple[ChunkEmbeddingResult, ChunkBatchStore], None, None]:
+    """Embed chunks to disk and yield a ``(result, store)`` pair.
+
+    The store owns the temp directory — files are cleaned up when the context
+    manager exits.
+
+    Usage::
+
+        with embed_and_stream(chunks, embedder, tenant_id, req_id) as (result, store):
+            for chunk in store.stream():
+                ...
+    """
+    with ChunkBatchStore() as store:
+        result = _embed_chunks_to_store(
+            chunks=chunks,
+            embedder=embedder,
+            tenant_id=tenant_id,
+            request_id=request_id,
+            store=store,
+        )
+        yield result, store
+
+
 def get_doc_ids_to_update(
     documents: list[Document], db_docs: list[DBDocument]
 ) -> list[Document]:
@@ -637,6 +762,29 @@ def add_contextual_summaries(
     return chunks
 
 
+def _verify_indexing_completeness(
+    insertion_records: list[DocumentInsertionRecord],
+    write_failures: list[ConnectorFailure],
+    embedding_failed_doc_ids: set[str],
+    updatable_ids: list[str],
+    document_index_name: str,
+) -> None:
+    """Verify that every updatable document was either indexed or reported as failed."""
+    all_returned_doc_ids = (
+        {r.document_id for r in insertion_records}
+        | {f.failed_document.document_id for f in write_failures if f.failed_document}
+        | embedding_failed_doc_ids
+    )
+    if all_returned_doc_ids != set(updatable_ids):
+        raise RuntimeError(
+            f"Some documents were not successfully indexed. "
+            f"Updatable IDs: {updatable_ids}, "
+            f"Returned IDs: {all_returned_doc_ids}. "
+            f"This should never happen. "
+            f"This occured for document index {document_index_name}"
+        )
+
+
 @log_function_time(debug_only=True)
 def index_doc_batch(
     *,
@@ -672,12 +820,7 @@ def index_doc_batch(
     filtered_documents = filter_fnc(document_batch)
     context = adapter.prepare(filtered_documents, ignore_time_skip)
     if not context:
-        return IndexingPipelineResult(
-            new_docs=0,
-            total_docs=len(filtered_documents),
-            total_chunks=0,
-            failures=[],
-        )
+        return IndexingPipelineResult.empty(len(filtered_documents))
 
     # Convert documents to IndexingDocument objects with processed section
     # logger.debug("Processing image sections")
@@ -716,117 +859,99 @@ def index_doc_batch(
         )
 
     logger.debug("Starting embedding")
-    chunks_with_embeddings, embedding_failures = (
-        embed_chunks_with_failure_handling(
-            chunks=chunks,
-            embedder=embedder,
-            tenant_id=tenant_id,
-            request_id=request_id,
-        )
-        if chunks
-        else ([], [])
-    )
+    with embed_and_stream(chunks, embedder, tenant_id, request_id) as (
+        embedding_result,
+        chunk_store,
+    ):
+        updatable_ids = [doc.id for doc in context.updatable_docs]
+        updatable_chunk_data = [
+            UpdatableChunkData(
+                chunk_id=chunk_id,
+                document_id=document_id,
+                boost_score=1.0,
+            )
+            for chunk_id, document_id in embedding_result.successful_chunk_ids
+        ]
 
-    chunk_content_scores = [1.0] * len(chunks_with_embeddings)
-
-    updatable_ids = [doc.id for doc in context.updatable_docs]
-    updatable_chunk_data = [
-        UpdatableChunkData(
-            chunk_id=chunk.chunk_id,
-            document_id=chunk.source_document.id,
-            boost_score=score,
-        )
-        for chunk, score in zip(chunks_with_embeddings, chunk_content_scores)
-    ]
-
-    # Acquires a lock on the documents so that no other process can modify them
-    # NOTE: don't need to acquire till here, since this is when the actual race condition
-    # with Vespa can occur.
-    with adapter.lock_context(context.updatable_docs):
-        # we're concerned about race conditions where multiple simultaneous indexings might result
-        # in one set of metadata overwriting another one in vespa.
-        # we still write data here for the immediate and most likely correct sync, but
-        # to resolve this, an update of the last modified field at the end of this loop
-        # always triggers a final metadata sync via the celery queue
-        result = adapter.build_metadata_aware_chunks(
-            chunks_with_embeddings=chunks_with_embeddings,
-            chunk_content_scores=chunk_content_scores,
-            tenant_id=tenant_id,
-            context=context,
+        embedding_failed_doc_ids = _get_failed_doc_ids(
+            embedding_result.connector_failures
         )
 
-        short_descriptor_list = [chunk.to_short_descriptor() for chunk in result.chunks]
-        short_descriptor_log = str(short_descriptor_list)[:1024]
-        logger.debug(f"Indexing the following chunks: {short_descriptor_log}")
+        # Filter to only successfully embedded chunks so
+        # doc_id_to_new_chunk_cnt reflects what's actually written to Vespa.
+        embedded_chunks = [
+            c for c in chunks if c.source_document.id not in embedding_failed_doc_ids
+        ]
 
-        primary_doc_idx_insertion_records: list[DocumentInsertionRecord] | None = None
-        primary_doc_idx_vector_db_write_failures: list[ConnectorFailure] | None = None
-        for document_index in document_indices:
-            # A document will not be spread across different batches, so all the
-            # documents with chunks in this set, are fully represented by the chunks
-            # in this set
-            (
-                insertion_records,
-                vector_db_write_failures,
-            ) = write_chunks_to_vector_db_with_backoff(
-                document_index=document_index,
-                chunks=result.chunks,
-                index_batch_params=IndexBatchParams(
-                    doc_id_to_previous_chunk_cnt=result.doc_id_to_previous_chunk_cnt,
-                    doc_id_to_new_chunk_cnt=result.doc_id_to_new_chunk_cnt,
-                    tenant_id=tenant_id,
-                    large_chunks_enabled=chunker.enable_large_chunks,
-                ),
+        # Acquires a lock on the documents so that no other process can modify
+        # them.  Not needed until here, since this is when the actual race
+        # condition with vector db can occur.
+        with adapter.lock_context(context.updatable_docs):
+            enricher = adapter.prepare_enrichment(
+                context=context,
+                tenant_id=tenant_id,
+                chunks=embedded_chunks,
             )
 
-            all_returned_doc_ids: set[str] = (
-                {record.document_id for record in insertion_records}
-                .union(
-                    {
-                        record.failed_document.document_id
-                        for record in vector_db_write_failures
-                        if record.failed_document
-                    }
-                )
-                .union(
-                    {
-                        record.failed_document.document_id
-                        for record in embedding_failures
-                        if record.failed_document
-                    }
-                )
+            index_batch_params = IndexBatchParams(
+                doc_id_to_previous_chunk_cnt=enricher.doc_id_to_previous_chunk_cnt,
+                doc_id_to_new_chunk_cnt=enricher.doc_id_to_new_chunk_cnt,
+                tenant_id=tenant_id,
+                large_chunks_enabled=chunker.enable_large_chunks,
             )
-            if all_returned_doc_ids != set(updatable_ids):
-                raise RuntimeError(
-                    f"Some documents were not successfully indexed. "
-                    f"Updatable IDs: {updatable_ids}, "
-                    f"Returned IDs: {all_returned_doc_ids}. "
-                    "This should never happen."
-                    f"This occured for document index {document_index.__class__.__name__}"
-                )
-            # We treat the first document index we got as the primary one used
-            # for reporting the state of indexing.
-            if primary_doc_idx_insertion_records is None:
-                primary_doc_idx_insertion_records = insertion_records
-            if primary_doc_idx_vector_db_write_failures is None:
-                primary_doc_idx_vector_db_write_failures = vector_db_write_failures
 
-        adapter.post_index(
-            context=context,
-            updatable_chunk_data=updatable_chunk_data,
-            filtered_documents=filtered_documents,
-            result=result,
-        )
+            primary_doc_idx_insertion_records: list[DocumentInsertionRecord] | None = (
+                None
+            )
+            primary_doc_idx_vector_db_write_failures: list[ConnectorFailure] | None = (
+                None
+            )
+
+            for document_index in document_indices:
+
+                def _enriched_stream() -> Iterator[DocMetadataAwareIndexChunk]:
+                    for chunk in chunk_store.stream():
+                        yield enricher.enrich_chunk(chunk, 1.0)
+
+                insertion_records, write_failures = (
+                    write_chunks_to_vector_db_with_backoff(
+                        document_index=document_index,
+                        make_chunks=_enriched_stream,
+                        index_batch_params=index_batch_params,
+                    )
+                )
+
+                _verify_indexing_completeness(
+                    insertion_records=insertion_records,
+                    write_failures=write_failures,
+                    embedding_failed_doc_ids=embedding_failed_doc_ids,
+                    updatable_ids=updatable_ids,
+                    document_index_name=document_index.__class__.__name__,
+                )
+                # We treat the first document index we got as the primary one used
+                # for reporting the state of indexing.
+                if primary_doc_idx_insertion_records is None:
+                    primary_doc_idx_insertion_records = insertion_records
+                if primary_doc_idx_vector_db_write_failures is None:
+                    primary_doc_idx_vector_db_write_failures = write_failures
+
+            adapter.post_index(
+                context=context,
+                updatable_chunk_data=updatable_chunk_data,
+                filtered_documents=filtered_documents,
+                enrichment=enricher,
+            )
 
     assert primary_doc_idx_insertion_records is not None
     assert primary_doc_idx_vector_db_write_failures is not None
     return IndexingPipelineResult(
-        new_docs=len(
-            [r for r in primary_doc_idx_insertion_records if not r.already_existed]
+        new_docs=sum(
+            1 for r in primary_doc_idx_insertion_records if not r.already_existed
         ),
         total_docs=len(filtered_documents),
-        total_chunks=len(chunks_with_embeddings),
-        failures=primary_doc_idx_vector_db_write_failures + embedding_failures,
+        total_chunks=len(embedding_result.successful_chunk_ids),
+        failures=primary_doc_idx_vector_db_write_failures
+        + embedding_result.connector_failures,
     )
 
 

backend/onyx/indexing/models.py

@@ -235,12 +235,16 @@ class UpdatableChunkData(BaseModel):
     boost_score: float
 
 
-class BuildMetadataAwareChunksResult(BaseModel):
-    chunks: list[DocMetadataAwareIndexChunk]
+class ChunkEnrichmentContext(Protocol):
+    """Returned by prepare_enrichment. Holds pre-computed metadata lookups
+    and provides per-chunk enrichment."""
+
     doc_id_to_previous_chunk_cnt: dict[str, int]
     doc_id_to_new_chunk_cnt: dict[str, int]
-    user_file_id_to_raw_text: dict[str, str]
-    user_file_id_to_token_count: dict[str, int | None]
+
+    def enrich_chunk(
+        self, chunk: IndexChunk, score: float
+    ) -> DocMetadataAwareIndexChunk: ...
 
 
 class IndexingBatchAdapter(Protocol):
@@ -254,18 +258,24 @@ class IndexingBatchAdapter(Protocol):
     ) -> Generator[TransactionalContext, None, None]:
         """Provide a transaction/row-lock context for critical updates."""
 
-    def build_metadata_aware_chunks(
+    def prepare_enrichment(
         self,
-        chunks_with_embeddings: list[IndexChunk],
-        chunk_content_scores: list[float],
-        tenant_id: str,
         context: "DocumentBatchPrepareContext",
-    ) -> BuildMetadataAwareChunksResult: ...
+        tenant_id: str,
+        chunks: list[DocAwareChunk],
+    ) -> ChunkEnrichmentContext:
+        """Prepare per-chunk enrichment data (access, document sets, boost, etc.).
+
+        Precondition: ``chunks`` have already been through the embedding step
+        (i.e. they are ``IndexChunk`` instances with populated embeddings,
+        passed here as the base ``DocAwareChunk`` type).
+        """
+        ...
 
     def post_index(
         self,
         context: "DocumentBatchPrepareContext",
         updatable_chunk_data: list[UpdatableChunkData],
         filtered_documents: list[Document],
-        result: BuildMetadataAwareChunksResult,
+        enrichment: ChunkEnrichmentContext,
     ) -> None: ...

backend/onyx/indexing/vector_db_insertion.py

@@ -1,6 +1,9 @@
 import time
-from collections import defaultdict
+from collections.abc import Callable
+from collections.abc import Iterable
 from http import HTTPStatus
+from itertools import chain
+from itertools import groupby
 
 import httpx
 
@@ -28,22 +31,22 @@ def _log_insufficient_storage_error(e: Exception) -> None:
 
 def write_chunks_to_vector_db_with_backoff(
     document_index: DocumentIndex,
-    chunks: list[DocMetadataAwareIndexChunk],
+    make_chunks: Callable[[], Iterable[DocMetadataAwareIndexChunk]],
     index_batch_params: IndexBatchParams,
 ) -> tuple[list[DocumentInsertionRecord], list[ConnectorFailure]]:
     """Tries to insert all chunks in one large batch. If that batch fails for any reason,
     goes document by document to isolate the failure(s).
 
     IMPORTANT: must pass in whole documents at a time not individual chunks, since the
-    vector DB interface assumes that all chunks for a single document are present.
+    vector DB interface assumes that all chunks for a single document are present. The
+    chunks must also be in contiguous batches
     """
-
     # first try to write the chunks to the vector db
     try:
         return (
             list(
                 document_index.index(
-                    chunks=chunks,
+                    chunks=make_chunks(),
                     index_batch_params=index_batch_params,
                 )
             ),
@@ -60,14 +63,23 @@ def write_chunks_to_vector_db_with_backoff(
         # wait a couple seconds just to give the vector db a chance to recover
         time.sleep(2)
 
-    # try writing each doc one by one
-    chunks_for_docs: dict[str, list[DocMetadataAwareIndexChunk]] = defaultdict(list)
-    for chunk in chunks:
-        chunks_for_docs[chunk.source_document.id].append(chunk)
-
     insertion_records: list[DocumentInsertionRecord] = []
     failures: list[ConnectorFailure] = []
-    for doc_id, chunks_for_doc in chunks_for_docs.items():
+
+    def key(chunk: DocMetadataAwareIndexChunk) -> str:
+        return chunk.source_document.id
+
+    seen_doc_ids: set[str] = set()
+    for doc_id, chunks_for_doc in groupby(make_chunks(), key=key):
+        if doc_id in seen_doc_ids:
+            raise RuntimeError(
+                f"Doc chunks are not arriving in order. Current doc_id={doc_id}, seen_doc_ids={list(seen_doc_ids)}"
+            )
+        seen_doc_ids.add(doc_id)
+
+        first_chunk = next(chunks_for_doc)
+        chunks_for_doc = chain([first_chunk], chunks_for_doc)
+
         try:
             insertion_records.extend(
                 document_index.index(
@@ -87,9 +99,7 @@ def write_chunks_to_vector_db_with_backoff(
                 ConnectorFailure(
                     failed_document=DocumentFailure(
                         document_id=doc_id,
-                        document_link=(
-                            chunks_for_doc[0].get_link() if chunks_for_doc else None
-                        ),
+                        document_link=first_chunk.get_link(),
                     ),
                     failure_message=str(e),
                     exception=e,

backend/onyx/llm/constants.py

@@ -25,6 +25,7 @@ class LlmProviderNames(str, Enum):
     LM_STUDIO = "lm_studio"
     MISTRAL = "mistral"
     LITELLM_PROXY = "litellm_proxy"
+    BIFROST = "bifrost"
 
     def __str__(self) -> str:
         """Needed so things like:
@@ -44,6 +45,7 @@ WELL_KNOWN_PROVIDER_NAMES = [
     LlmProviderNames.OLLAMA_CHAT,
     LlmProviderNames.LM_STUDIO,
     LlmProviderNames.LITELLM_PROXY,
+    LlmProviderNames.BIFROST,
 ]
 
 
@@ -61,6 +63,7 @@ PROVIDER_DISPLAY_NAMES: dict[str, str] = {
     LlmProviderNames.OLLAMA_CHAT: "Ollama",
     LlmProviderNames.LM_STUDIO: "LM Studio",
     LlmProviderNames.LITELLM_PROXY: "LiteLLM Proxy",
+    LlmProviderNames.BIFROST: "Bifrost",
     "groq": "Groq",
     "anyscale": "Anyscale",
     "deepseek": "DeepSeek",
@@ -112,6 +115,7 @@ AGGREGATOR_PROVIDERS: set[str] = {
     LlmProviderNames.VERTEX_AI,
     LlmProviderNames.AZURE,
     LlmProviderNames.LITELLM_PROXY,
+    LlmProviderNames.BIFROST,
 }
 
 # Model family name mappings for display name generation

backend/onyx/llm/multi_llm.py

@@ -185,6 +185,21 @@ def _messages_contain_tool_content(messages: list[dict[str, Any]]) -> bool:
     return False
 
 
+def _prompt_contains_tool_call_history(prompt: LanguageModelInput) -> bool:
+    """Check if the prompt contains any assistant messages with tool_calls.
+
+    When Anthropic's extended thinking is enabled, the API requires every
+    assistant message to start with a thinking block before any tool_use
+    blocks.  Since we don't preserve thinking_blocks (they carry
+    cryptographic signatures that can't be reconstructed), we must skip
+    the thinking param whenever history contains prior tool-calling turns.
+    """
+    from onyx.llm.models import AssistantMessage
+
+    msgs = prompt if isinstance(prompt, list) else [prompt]
+    return any(isinstance(msg, AssistantMessage) and msg.tool_calls for msg in msgs)
+
+
 def _is_vertex_model_rejecting_output_config(model_name: str) -> bool:
     normalized_model_name = model_name.lower()
     return any(
@@ -290,6 +305,17 @@ class LitellmLLM(LLM):
         ):
             model_kwargs[VERTEX_LOCATION_KWARG] = "global"
 
+        # Bifrost: OpenAI-compatible proxy that expects model names in
+        # provider/model format (e.g. "anthropic/claude-sonnet-4-6").
+        # We route through LiteLLM's openai provider with the Bifrost base URL,
+        # and ensure /v1 is appended.
+        if model_provider == LlmProviderNames.BIFROST:
+            self._custom_llm_provider = "openai"
+            if self._api_base is not None:
+                base = self._api_base.rstrip("/")
+                self._api_base = base if base.endswith("/v1") else f"{base}/v1"
+                model_kwargs["api_base"] = self._api_base
+
         # This is needed for Ollama to do proper function calling
         if model_provider == LlmProviderNames.OLLAMA_CHAT and api_base is not None:
             model_kwargs["api_base"] = api_base
@@ -401,14 +427,20 @@ class LitellmLLM(LLM):
         optional_kwargs: dict[str, Any] = {}
 
         # Model name
+        is_bifrost = self._model_provider == LlmProviderNames.BIFROST
         model_provider = (
             f"{self.config.model_provider}/responses"
             if is_openai_model  # Uses litellm's completions -> responses bridge
             else self.config.model_provider
         )
-        model = (
-            f"{model_provider}/{self.config.deployment_name or self.config.model_name}"
-        )
+        if is_bifrost:
+            # Bifrost expects model names in provider/model format
+            # (e.g. "anthropic/claude-sonnet-4-6") sent directly to its
+            # OpenAI-compatible endpoint. We use custom_llm_provider="openai"
+            # so LiteLLM doesn't try to route based on the provider prefix.
+            model = self.config.deployment_name or self.config.model_name
+        else:
+            model = f"{model_provider}/{self.config.deployment_name or self.config.model_name}"
 
         # Tool choice
         if is_claude_model and tool_choice == ToolChoiceOptions.REQUIRED:
@@ -449,7 +481,20 @@ class LitellmLLM(LLM):
                     reasoning_effort
                 )
 
-                if budget_tokens is not None:
+                # Anthropic requires every assistant message with tool_use
+                # blocks to start with a thinking block that carries a
+                # cryptographic signature.  We don't preserve those blocks
+                # across turns, so skip thinking when the history already
+                # contains tool-calling assistant messages.  LiteLLM's
+                # modify_params workaround doesn't cover all providers
+                # (notably Bedrock).
+                can_enable_thinking = (
+                    budget_tokens is not None
+                    and not _prompt_contains_tool_call_history(prompt)
+                )
+
+                if can_enable_thinking:
+                    assert budget_tokens is not None  # mypy
                     if max_tokens is not None:
                         # Anthropic has a weird rule where max token has to be at least as much as budget tokens if set
                         # and the minimum budget tokens is 1024
@@ -483,10 +528,11 @@ class LitellmLLM(LLM):
         if structured_response_format:
             optional_kwargs["response_format"] = structured_response_format
 
-        if not (is_claude_model or is_ollama or is_mistral):
+        if not (is_claude_model or is_ollama or is_mistral) or is_bifrost:
             # Litellm bug: tool_choice is dropped silently if not specified here for OpenAI
             # However, this param breaks Anthropic and Mistral models,
-            # so it must be conditionally included.
+            # so it must be conditionally included unless the request is
+            # routed through Bifrost's OpenAI-compatible endpoint.
             # Additionally, tool_choice is not supported by Ollama and causes warnings if included.
             # See also, https://github.com/ollama/ollama/issues/11171
             optional_kwargs["allowed_openai_params"] = ["tool_choice"]

backend/onyx/llm/override_models.py

@@ -8,6 +8,24 @@ from pydantic import BaseModel
 
 
 class LLMOverride(BaseModel):
+    """Per-request LLM settings that override persona defaults.
+
+    All fields are optional — only the fields that differ from the persona's
+    configured LLM need to be supplied. Used both over the wire (API requests)
+    and for multi-model comparison, where one override is supplied per model.
+
+    Attributes:
+        model_provider: LLM provider slug (e.g. ``"openai"``, ``"anthropic"``).
+            When ``None``, the persona's default provider is used.
+        model_version: Specific model version string (e.g. ``"gpt-4o"``).
+            When ``None``, the persona's default model is used.
+        temperature: Sampling temperature in ``[0, 2]``. When ``None``, the
+            persona's default temperature is used.
+        display_name: Human-readable label shown in the UI for this model,
+            e.g. ``"GPT-4 Turbo"``. Optional; falls back to ``model_version``
+            when not set.
+    """
+
     model_provider: str | None = None
     model_version: str | None = None
     temperature: float | None = None

backend/onyx/llm/well_known_providers/constants.py

@@ -13,6 +13,8 @@ LM_STUDIO_API_KEY_CONFIG_KEY = "LM_STUDIO_API_KEY"
 
 LITELLM_PROXY_PROVIDER_NAME = "litellm_proxy"
 
+BIFROST_PROVIDER_NAME = "bifrost"
+
 # Providers that use optional Bearer auth from custom_config
 PROVIDERS_WITH_SPECIAL_API_KEY_HANDLING: dict[str, str] = {
     LlmProviderNames.OLLAMA_CHAT: OLLAMA_API_KEY_CONFIG_KEY,

backend/onyx/llm/well_known_providers/llm_provider_options.py

@@ -15,6 +15,7 @@ from onyx.llm.well_known_providers.auto_update_service import (
 from onyx.llm.well_known_providers.constants import ANTHROPIC_PROVIDER_NAME
 from onyx.llm.well_known_providers.constants import AZURE_PROVIDER_NAME
 from onyx.llm.well_known_providers.constants import BEDROCK_PROVIDER_NAME
+from onyx.llm.well_known_providers.constants import BIFROST_PROVIDER_NAME
 from onyx.llm.well_known_providers.constants import LITELLM_PROXY_PROVIDER_NAME
 from onyx.llm.well_known_providers.constants import LM_STUDIO_PROVIDER_NAME
 from onyx.llm.well_known_providers.constants import OLLAMA_PROVIDER_NAME
@@ -49,6 +50,7 @@ def _get_provider_to_models_map() -> dict[str, list[str]]:
         LM_STUDIO_PROVIDER_NAME: [],  # Dynamic - fetched from LM Studio API
         OPENROUTER_PROVIDER_NAME: [],  # Dynamic - fetched from OpenRouter API
         LITELLM_PROXY_PROVIDER_NAME: [],  # Dynamic - fetched from LiteLLM proxy API
+        BIFROST_PROVIDER_NAME: [],  # Dynamic - fetched from Bifrost API
     }
 
 

backend/onyx/natural_language_processing/utils.py

@@ -175,6 +175,32 @@ def get_tokenizer(
     return _check_tokenizer_cache(provider_type, model_name)
 
 
+# Max characters per encode() call.
+_ENCODE_CHUNK_SIZE = 500_000
+
+
+def count_tokens(
+    text: str,
+    tokenizer: BaseTokenizer,
+    token_limit: int | None = None,
+) -> int:
+    """Count tokens, chunking the input to avoid tiktoken stack overflow.
+
+    If token_limit is provided and the text is large enough to require
+    multiple chunks (> 500k chars), stops early once the count exceeds it.
+    When early-exiting, the returned value exceeds token_limit but may be
+    less than the true full token count.
+    """
+    if len(text) <= _ENCODE_CHUNK_SIZE:
+        return len(tokenizer.encode(text))
+    total = 0
+    for start in range(0, len(text), _ENCODE_CHUNK_SIZE):
+        total += len(tokenizer.encode(text[start : start + _ENCODE_CHUNK_SIZE]))
+        if token_limit is not None and total > token_limit:
+            return total  # Already over — skip remaining chunks
+    return total
+
+
 def tokenizer_trim_content(
     content: str, desired_length: int, tokenizer: BaseTokenizer
 ) -> str:

backend/onyx/server/features/build/sandbox/kubernetes/docker/templates/outputs/web/package-lock.json

@@ -3844,9 +3844,9 @@
       }
     },
     "node_modules/@ts-morph/common/node_modules/brace-expansion": {
-      "version": "5.0.3",
-      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-5.0.3.tgz",
-      "integrity": "sha512-fy6KJm2RawA5RcHkLa1z/ScpBeA762UF9KmZQxwIbDtRJrgLzM10depAiEQ+CXYcoiqW1/m96OAAoke2nE9EeA==",
+      "version": "5.0.5",
+      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-5.0.5.tgz",
+      "integrity": "sha512-VZznLgtwhn+Mact9tfiwx64fA9erHH/MCXEUfB/0bX/6Fz6ny5EGTXYltMocqg4xFAQZtnO3DHWWXi8RiuN7cQ==",
       "license": "MIT",
       "dependencies": {
         "balanced-match": "^4.0.2"
@@ -4224,9 +4224,9 @@
       }
     },
     "node_modules/@typescript-eslint/typescript-estree/node_modules/brace-expansion": {
-      "version": "2.0.2",
-      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-2.0.2.tgz",
-      "integrity": "sha512-Jt0vHyM+jmUBqojB7E1NIYadt0vI0Qxjxd2TErW94wDz+E2LAm5vKMXXwg6ZZBTHPuUlDgQHKXvjGBdfcF1ZDQ==",
+      "version": "2.0.3",
+      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-2.0.3.tgz",
+      "integrity": "sha512-MCV/fYJEbqx68aE58kv2cA/kiky1G8vux3OR6/jbS+jIMe/6fJWa0DTzJU7dqijOWYwHi1t29FlfYI9uytqlpA==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
@@ -5007,9 +5007,9 @@
       }
     },
     "node_modules/brace-expansion": {
-      "version": "1.1.12",
-      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-1.1.12.tgz",
-      "integrity": "sha512-9T9UjW3r0UW5c1Q7GTwllptXwhvYmEzFhzMfZ9H7FQWt+uZePjZPjBP/W1ZEyZ1twGWom5/56TF4lPcqjnDHcg==",
+      "version": "1.1.13",
+      "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-1.1.13.tgz",
+      "integrity": "sha512-9ZLprWS6EENmhEOpjCYW2c8VkmOvckIJZfkr7rBW6dObmfgJ/L1GpSYW5Hpo9lDz4D1+n0Ckz8rU7FwHDQiG/w==",
       "dev": true,
       "license": "MIT",
       "dependencies": {

backend/onyx/server/features/hooks/api.py

@@ -44,11 +44,12 @@ def _check_ssrf_safety(endpoint_url: str) -> None:
     """Raise OnyxError if endpoint_url could be used for SSRF.
 
     Delegates to validate_outbound_http_url with https_only=True.
+    Uses BAD_GATEWAY so the frontend maps the error to the Endpoint URL field.
     """
     try:
         validate_outbound_http_url(endpoint_url, https_only=True)
     except (SSRFException, ValueError) as e:
-        raise OnyxError(OnyxErrorCode.INVALID_INPUT, str(e))
+        raise OnyxError(OnyxErrorCode.BAD_GATEWAY, str(e))
 
 
 # ---------------------------------------------------------------------------
@@ -122,9 +123,8 @@ def _validate_endpoint(
     (not reachable — indicates the api_key is invalid).
 
     Timeout handling:
-    - ConnectTimeout: TCP handshake never completed → cannot_connect.
-    - ReadTimeout / WriteTimeout: TCP was established, server responded slowly → timeout
-      (operator should consider increasing timeout_seconds).
+    - Any httpx.TimeoutException (ConnectTimeout, ReadTimeout, WriteTimeout, PoolTimeout) →
+      timeout (operator should consider increasing timeout_seconds).
     - All other exceptions → cannot_connect.
     """
     _check_ssrf_safety(endpoint_url)
@@ -141,19 +141,11 @@ def _validate_endpoint(
             )
         return HookValidateResponse(status=HookValidateStatus.passed)
     except httpx.TimeoutException as exc:
-        # ConnectTimeout: TCP handshake never completed → cannot_connect.
-        # ReadTimeout / WriteTimeout: TCP was established, server just responded slowly → timeout.
-        if isinstance(exc, httpx.ConnectTimeout):
-            logger.warning(
-                "Hook endpoint validation: connect timeout for %s",
-                endpoint_url,
-                exc_info=exc,
-            )
-            return HookValidateResponse(
-                status=HookValidateStatus.cannot_connect, error_message=str(exc)
-            )
+        # Any timeout (connect, read, or write) means the configured timeout_seconds
+        # is too low for this endpoint. Report as timeout so the UI directs the user
+        # to increase the timeout setting.
         logger.warning(
-            "Hook endpoint validation: read/write timeout for %s",
+            "Hook endpoint validation: timeout for %s",
             endpoint_url,
             exc_info=exc,
         )

backend/onyx/server/features/projects/projects_file_utils.py

@@ -9,20 +9,15 @@ from pydantic import ConfigDict
 from pydantic import Field
 from sqlalchemy.orm import Session
 
-from onyx.configs.app_configs import FILE_TOKEN_COUNT_THRESHOLD
-from onyx.configs.app_configs import USER_FILE_MAX_UPLOAD_SIZE_BYTES
-from onyx.configs.app_configs import USER_FILE_MAX_UPLOAD_SIZE_MB
 from onyx.db.llm import fetch_default_llm_model
 from onyx.file_processing.extract_file_text import extract_file_text
 from onyx.file_processing.extract_file_text import get_file_ext
 from onyx.file_processing.file_types import OnyxFileExtensions
 from onyx.file_processing.password_validation import is_file_password_protected
+from onyx.natural_language_processing.utils import count_tokens
 from onyx.natural_language_processing.utils import get_tokenizer
+from onyx.server.settings.store import load_settings
 from onyx.utils.logger import setup_logger
-from shared_configs.configs import MULTI_TENANT
-from shared_configs.configs import SKIP_USERFILE_THRESHOLD
-from shared_configs.configs import SKIP_USERFILE_THRESHOLD_TENANT_LIST
-from shared_configs.contextvars import get_current_tenant_id
 
 
 logger = setup_logger()
@@ -161,8 +156,8 @@ def categorize_uploaded_files(
       document formats (.pdf, .docx, …) and falls back to a text-detection
       heuristic for unknown extensions (.py, .js, .rs, …).
     - Uses default tokenizer to compute token length.
-    - If token length > threshold, reject file (unless threshold skip is enabled).
-    - If text cannot be extracted, reject file.
+    - If token length exceeds the admin-configured threshold, reject file.
+    - If extension unsupported or text cannot be extracted, reject file.
     - Otherwise marked as acceptable.
     """
 
@@ -173,36 +168,33 @@ def categorize_uploaded_files(
     provider_type = default_model.llm_provider.provider if default_model else None
     tokenizer = get_tokenizer(model_name=model_name, provider_type=provider_type)
 
-    # Check if threshold checks should be skipped
-    skip_threshold = False
-
-    # Check global skip flag (works for both single-tenant and multi-tenant)
-    if SKIP_USERFILE_THRESHOLD:
-        skip_threshold = True
-        logger.info("Skipping userfile threshold check (global setting)")
-    # Check tenant-specific skip list (only applicable in multi-tenant)
-    elif MULTI_TENANT and SKIP_USERFILE_THRESHOLD_TENANT_LIST:
-        try:
-            current_tenant_id = get_current_tenant_id()
-            skip_threshold = current_tenant_id in SKIP_USERFILE_THRESHOLD_TENANT_LIST
-            if skip_threshold:
-                logger.info(
-                    f"Skipping userfile threshold check for tenant: {current_tenant_id}"
-                )
-        except RuntimeError as e:
-            logger.warning(f"Failed to get current tenant ID: {str(e)}")
+    # Derive limits from admin-configurable settings.
+    # For upload size: load_settings() resolves 0/None to a positive default.
+    # For token threshold: 0 means "no limit" (converted to None below).
+    settings = load_settings()
+    max_upload_size_mb = (
+        settings.user_file_max_upload_size_mb
+    )  # always positive after load_settings()
+    max_upload_size_bytes = (
+        max_upload_size_mb * 1024 * 1024 if max_upload_size_mb else None
+    )
+    token_threshold_k = settings.file_token_count_threshold_k
+    token_threshold = (
+        token_threshold_k * 1000 if token_threshold_k else None
+    )  # 0 → None = no limit
 
     for upload in files:
         try:
             filename = get_safe_filename(upload)
 
-            # Size limit is a hard safety cap and is enforced even when token
-            # threshold checks are skipped via SKIP_USERFILE_THRESHOLD settings.
-            if is_upload_too_large(upload, USER_FILE_MAX_UPLOAD_SIZE_BYTES):
+            # Size limit is a hard safety cap.
+            if max_upload_size_bytes is not None and is_upload_too_large(
+                upload, max_upload_size_bytes
+            ):
                 results.rejected.append(
                     RejectedFile(
                         filename=filename,
-                        reason=f"Exceeds {USER_FILE_MAX_UPLOAD_SIZE_MB} MB file size limit",
+                        reason=f"Exceeds {max_upload_size_mb} MB file size limit",
                     )
                 )
                 continue
@@ -224,11 +216,11 @@ def categorize_uploaded_files(
                     )
                     continue
 
-                if not skip_threshold and token_count > FILE_TOKEN_COUNT_THRESHOLD:
+                if token_threshold is not None and token_count > token_threshold:
                     results.rejected.append(
                         RejectedFile(
                             filename=filename,
-                            reason=f"Exceeds {FILE_TOKEN_COUNT_THRESHOLD} token limit",
+                            reason=f"Exceeds {token_threshold_k}K token limit",
                         )
                     )
                 else:
@@ -269,12 +261,14 @@ def categorize_uploaded_files(
                     )
                     continue
 
-                token_count = len(tokenizer.encode(text_content))
-                if not skip_threshold and token_count > FILE_TOKEN_COUNT_THRESHOLD:
+                token_count = count_tokens(
+                    text_content, tokenizer, token_limit=token_threshold
+                )
+                if token_threshold is not None and token_count > token_threshold:
                     results.rejected.append(
                         RejectedFile(
                             filename=filename,
-                            reason=f"Exceeds {FILE_TOKEN_COUNT_THRESHOLD} token limit",
+                            reason=f"Exceeds {token_threshold_k}K token limit",
                         )
                     )
                 else:

backend/onyx/server/manage/llm/api.py

@@ -57,6 +57,8 @@ from onyx.llm.well_known_providers.llm_provider_options import (
 )
 from onyx.server.manage.llm.models import BedrockFinalModelResponse
 from onyx.server.manage.llm.models import BedrockModelsRequest
+from onyx.server.manage.llm.models import BifrostFinalModelResponse
+from onyx.server.manage.llm.models import BifrostModelsRequest
 from onyx.server.manage.llm.models import DefaultModel
 from onyx.server.manage.llm.models import LitellmFinalModelResponse
 from onyx.server.manage.llm.models import LitellmModelDetails
@@ -1422,11 +1424,26 @@ def _get_litellm_models_response(api_key: str, api_base: str) -> dict:
     cleaned_api_base = api_base.strip().rstrip("/")
     url = f"{cleaned_api_base}/v1/models"
 
+    return _get_openai_compatible_models_response(
+        url=url,
+        source_name="LiteLLM proxy",
+        api_key=api_key,
+    )
+
+
+def _get_openai_compatible_models_response(
+    url: str,
+    source_name: str,
+    api_key: str | None = None,
+) -> dict:
+    """Fetch model metadata from an OpenAI-compatible `/models` endpoint."""
     headers = {
         "Authorization": f"Bearer {api_key}",
         "HTTP-Referer": "https://onyx.app",
         "X-Title": "Onyx",
     }
+    if not api_key:
+        headers.pop("Authorization")
 
     try:
         response = httpx.get(url, headers=headers, timeout=10.0)
@@ -1436,20 +1453,125 @@ def _get_litellm_models_response(api_key: str, api_base: str) -> dict:
         if e.response.status_code == 401:
             raise OnyxError(
                 OnyxErrorCode.VALIDATION_ERROR,
-                "Authentication failed: invalid or missing API key for LiteLLM proxy.",
+                f"Authentication failed: invalid or missing API key for {source_name}.",
             )
         elif e.response.status_code == 404:
             raise OnyxError(
                 OnyxErrorCode.VALIDATION_ERROR,
-                f"LiteLLM models endpoint not found at {url}. Please verify the API base URL.",
+                f"{source_name} models endpoint not found at {url}. Please verify the API base URL.",
             )
         else:
             raise OnyxError(
                 OnyxErrorCode.BAD_GATEWAY,
-                f"Failed to fetch LiteLLM models: {e}",
+                f"Failed to fetch {source_name} models: {e}",
             )
-    except Exception as e:
+    except httpx.RequestError as e:
+        logger.warning(
+            "Failed to fetch models from OpenAI-compatible endpoint",
+            extra={"source": source_name, "url": url, "error": str(e)},
+            exc_info=True,
+        )
         raise OnyxError(
             OnyxErrorCode.BAD_GATEWAY,
-            f"Failed to fetch LiteLLM models: {e}",
+            f"Failed to fetch {source_name} models: {e}",
         )
+    except ValueError as e:
+        logger.warning(
+            "Received invalid model response from OpenAI-compatible endpoint",
+            extra={"source": source_name, "url": url, "error": str(e)},
+            exc_info=True,
+        )
+        raise OnyxError(
+            OnyxErrorCode.BAD_GATEWAY,
+            f"Failed to fetch {source_name} models: {e}",
+        )
+
+
+@admin_router.post("/bifrost/available-models")
+def get_bifrost_available_models(
+    request: BifrostModelsRequest,
+    _: User = Depends(current_admin_user),
+    db_session: Session = Depends(get_session),
+) -> list[BifrostFinalModelResponse]:
+    """Fetch available models from Bifrost gateway /v1/models endpoint."""
+    response_json = _get_bifrost_models_response(
+        api_base=request.api_base, api_key=request.api_key
+    )
+
+    models = response_json.get("data", [])
+    if not isinstance(models, list) or len(models) == 0:
+        raise OnyxError(
+            OnyxErrorCode.VALIDATION_ERROR,
+            "No models found from your Bifrost endpoint",
+        )
+
+    results: list[BifrostFinalModelResponse] = []
+    for model in models:
+        try:
+            model_id = model.get("id", "")
+            model_name = model.get("name", model_id)
+
+            if not model_id:
+                continue
+
+            # Skip embedding models
+            if is_embedding_model(model_id):
+                continue
+
+            results.append(
+                BifrostFinalModelResponse(
+                    name=model_id,
+                    display_name=model_name,
+                    max_input_tokens=model.get("context_length"),
+                    supports_image_input=infer_vision_support(model_id),
+                    supports_reasoning=is_reasoning_model(model_id, model_name),
+                )
+            )
+        except Exception as e:
+            logger.warning(
+                "Failed to parse Bifrost model entry",
+                extra={"error": str(e), "item": str(model)[:1000]},
+            )
+
+    if not results:
+        raise OnyxError(
+            OnyxErrorCode.VALIDATION_ERROR,
+            "No compatible models found from Bifrost",
+        )
+
+    sorted_results = sorted(results, key=lambda m: m.name.lower())
+
+    # Sync new models to DB if provider_name is specified
+    if request.provider_name:
+        _sync_fetched_models(
+            db_session=db_session,
+            provider_name=request.provider_name,
+            models=[
+                SyncModelEntry(
+                    name=r.name,
+                    display_name=r.display_name,
+                    max_input_tokens=r.max_input_tokens,
+                    supports_image_input=r.supports_image_input,
+                )
+                for r in sorted_results
+            ],
+            source_label="Bifrost",
+        )
+
+    return sorted_results
+
+
+def _get_bifrost_models_response(api_base: str, api_key: str | None = None) -> dict:
+    """Perform GET to Bifrost /v1/models and return parsed JSON."""
+    cleaned_api_base = api_base.strip().rstrip("/")
+    # Ensure we hit /v1/models
+    if cleaned_api_base.endswith("/v1"):
+        url = f"{cleaned_api_base}/models"
+    else:
+        url = f"{cleaned_api_base}/v1/models"
+
+    return _get_openai_compatible_models_response(
+        url=url,
+        source_name="Bifrost",
+        api_key=api_key,
+    )

backend/onyx/server/manage/llm/models.py

@@ -449,3 +449,18 @@ class LitellmModelDetails(BaseModel):
 class LitellmFinalModelResponse(BaseModel):
     provider_name: str  # Provider name (e.g. "openai")
     model_name: str  # Model ID (e.g. "gpt-4o")
+
+
+# Bifrost dynamic models fetch
+class BifrostModelsRequest(BaseModel):
+    api_base: str
+    api_key: str | None = None
+    provider_name: str | None = None  # Optional: to save models to existing provider
+
+
+class BifrostFinalModelResponse(BaseModel):
+    name: str  # Model ID in provider/model format (e.g. "anthropic/claude-sonnet-4-6")
+    display_name: str  # Human-readable name from Bifrost API
+    max_input_tokens: int | None
+    supports_image_input: bool
+    supports_reasoning: bool

backend/onyx/server/manage/llm/utils.py

@@ -25,6 +25,7 @@ DYNAMIC_LLM_PROVIDERS = frozenset(
         LlmProviderNames.BEDROCK,
         LlmProviderNames.OLLAMA_CHAT,
         LlmProviderNames.LM_STUDIO,
+        LlmProviderNames.BIFROST,
     }
 )
 
@@ -50,6 +51,25 @@ BEDROCK_VISION_MODELS = frozenset(
     }
 )
 
+# Known Bifrost/OpenAI-compatible vision-capable model families where the
+# source API does not expose this metadata directly.
+BIFROST_VISION_MODEL_FAMILIES = frozenset(
+    {
+        "anthropic/claude-3",
+        "anthropic/claude-4",
+        "amazon/nova-pro",
+        "amazon/nova-lite",
+        "amazon/nova-premier",
+        "openai/gpt-4o",
+        "openai/gpt-4.1",
+        "google/gemini",
+        "meta-llama/llama-3.2",
+        "mistral/pixtral",
+        "qwen/qwen2.5-vl",
+        "qwen/qwen-vl",
+    }
+)
+
 
 def is_valid_bedrock_model(
     model_id: str,
@@ -76,11 +96,18 @@ def is_valid_bedrock_model(
 def infer_vision_support(model_id: str) -> bool:
     """Infer vision support from model ID when base model metadata unavailable.
 
-    Used for cross-region inference profiles when the base model isn't
-    available in the user's region.
+    Used for providers like Bedrock and Bifrost where vision support may
+    need to be inferred from vendor/model naming conventions.
     """
     model_id_lower = model_id.lower()
-    return any(vision_model in model_id_lower for vision_model in BEDROCK_VISION_MODELS)
+    if any(vision_model in model_id_lower for vision_model in BEDROCK_VISION_MODELS):
+        return True
+
+    normalized_model_id = model_id_lower.replace(".", "/")
+    return any(
+        vision_model in normalized_model_id
+        for vision_model in BIFROST_VISION_MODEL_FAMILIES
+    )
 
 
 def generate_bedrock_display_name(model_id: str) -> str:
@@ -322,7 +349,7 @@ def extract_vendor_from_model_name(model_name: str, provider: str) -> str | None
         - Ollama: "llama3:70b" → "Meta"
         - Ollama: "qwen2.5:7b" → "Alibaba"
     """
-    if provider == LlmProviderNames.OPENROUTER:
+    if provider in (LlmProviderNames.OPENROUTER, LlmProviderNames.BIFROST):
         # Format: "vendor/model-name" e.g., "anthropic/claude-3-5-sonnet"
         if "/" in model_name:
             vendor_key = model_name.split("/")[0].lower()

backend/onyx/server/metrics/indexing_pipeline.py

@@ -12,7 +12,6 @@ stale, which is fine for monitoring dashboards.
 import json
 import threading
 import time
-from collections.abc import Callable
 from datetime import datetime
 from datetime import timezone
 from typing import Any
@@ -104,25 +103,23 @@ class _CachedCollector(Collector):
 
 
 class QueueDepthCollector(_CachedCollector):
-    """Reads Celery queue lengths from the broker Redis on each scrape.
-
-    Uses a Redis client factory (callable) rather than a stored client
-    reference so the connection is always fresh from Celery's pool.
-    """
+    """Reads Celery queue lengths from the broker Redis on each scrape."""
 
     def __init__(self, cache_ttl: float = _DEFAULT_CACHE_TTL) -> None:
         super().__init__(cache_ttl)
-        self._get_redis: Callable[[], Redis] | None = None
+        self._celery_app: Any | None = None
 
-    def set_redis_factory(self, factory: Callable[[], Redis]) -> None:
-        """Set a callable that returns a broker Redis client on demand."""
-        self._get_redis = factory
+    def set_celery_app(self, app: Any) -> None:
+        """Set the Celery app for broker Redis access."""
+        self._celery_app = app
 
     def _collect_fresh(self) -> list[GaugeMetricFamily]:
-        if self._get_redis is None:
+        if self._celery_app is None:
             return []
 
-        redis_client = self._get_redis()
+        from onyx.background.celery.celery_redis import celery_get_broker_client
+
+        redis_client = celery_get_broker_client(self._celery_app)
 
         depth = GaugeMetricFamily(
             "onyx_queue_depth",
@@ -404,17 +401,19 @@ class RedisHealthCollector(_CachedCollector):
 
     def __init__(self, cache_ttl: float = _DEFAULT_CACHE_TTL) -> None:
         super().__init__(cache_ttl)
-        self._get_redis: Callable[[], Redis] | None = None
+        self._celery_app: Any | None = None
 
-    def set_redis_factory(self, factory: Callable[[], Redis]) -> None:
-        """Set a callable that returns a broker Redis client on demand."""
-        self._get_redis = factory
+    def set_celery_app(self, app: Any) -> None:
+        """Set the Celery app for broker Redis access."""
+        self._celery_app = app
 
     def _collect_fresh(self) -> list[GaugeMetricFamily]:
-        if self._get_redis is None:
+        if self._celery_app is None:
             return []
 
-        redis_client = self._get_redis()
+        from onyx.background.celery.celery_redis import celery_get_broker_client
+
+        redis_client = celery_get_broker_client(self._celery_app)
 
         memory_used = GaugeMetricFamily(
             "onyx_redis_memory_used_bytes",
@@ -449,40 +448,128 @@ class RedisHealthCollector(_CachedCollector):
         return [memory_used, memory_peak, memory_frag, connected_clients]
 
 
-class WorkerHealthCollector(_CachedCollector):
-    """Collects Celery worker count and process count via inspect ping.
+class WorkerHeartbeatMonitor:
+    """Monitors Celery worker health via the event stream.
 
-    Uses a longer cache TTL (60s) since inspect.ping() is a broadcast
-    command that takes a couple seconds to complete.
-
-    Maintains a set of known worker short-names so that when a worker
-    stops responding, we emit ``up=0`` instead of silently dropping the
-    metric (which would make ``absent()``-style alerts impossible).
+    Subscribes to ``worker-heartbeat``, ``worker-online``, and
+    ``worker-offline`` events via a single persistent connection.
+    Runs in a daemon thread started once during worker setup.
     """
 
-    # Remove a worker from _known_workers after this many consecutive
-    # missed pings (at 60s TTL ≈ 10 minutes of being unreachable).
-    _MAX_CONSECUTIVE_MISSES = 10
+    # Consider a worker down if no heartbeat received for this long.
+    _HEARTBEAT_TIMEOUT_SECONDS = 120.0
 
-    def __init__(self, cache_ttl: float = 60.0) -> None:
+    def __init__(self, celery_app: Any) -> None:
+        self._app = celery_app
+        self._worker_last_seen: dict[str, float] = {}
+        self._lock = threading.Lock()
+        self._running = False
+        self._thread: threading.Thread | None = None
+
+    def start(self) -> None:
+        """Start the background event listener thread.
+
+        Safe to call multiple times — only starts one thread.
+        """
+        if self._thread is not None and self._thread.is_alive():
+            return
+        self._running = True
+        self._thread = threading.Thread(target=self._listen, daemon=True)
+        self._thread.start()
+        logger.info("WorkerHeartbeatMonitor started")
+
+    def stop(self) -> None:
+        self._running = False
+
+    def _listen(self) -> None:
+        """Background loop: connect to event stream and process heartbeats."""
+        while self._running:
+            try:
+                with self._app.connection() as conn:
+                    recv = self._app.events.Receiver(
+                        conn,
+                        handlers={
+                            "worker-heartbeat": self._on_heartbeat,
+                            "worker-online": self._on_heartbeat,
+                            "worker-offline": self._on_offline,
+                        },
+                    )
+                    recv.capture(
+                        limit=None, timeout=self._HEARTBEAT_TIMEOUT_SECONDS, wakeup=True
+                    )
+            except Exception:
+                if self._running:
+                    logger.debug(
+                        "Heartbeat listener disconnected, reconnecting in 5s",
+                        exc_info=True,
+                    )
+                    time.sleep(5.0)
+            else:
+                # capture() returned normally (timeout with no events); reconnect
+                if self._running:
+                    logger.debug("Heartbeat capture timed out, reconnecting")
+                    time.sleep(5.0)
+
+    def _on_heartbeat(self, event: dict[str, Any]) -> None:
+        hostname = event.get("hostname")
+        if hostname:
+            with self._lock:
+                self._worker_last_seen[hostname] = time.monotonic()
+
+    def _on_offline(self, event: dict[str, Any]) -> None:
+        hostname = event.get("hostname")
+        if hostname:
+            with self._lock:
+                self._worker_last_seen.pop(hostname, None)
+
+    def get_worker_status(self) -> dict[str, bool]:
+        """Return {hostname: is_alive} for all known workers.
+
+        Thread-safe. Called by WorkerHealthCollector on each scrape.
+        Also prunes workers that have been dead longer than 2x the
+        heartbeat timeout to prevent unbounded growth.
+        """
+        now = time.monotonic()
+        prune_threshold = self._HEARTBEAT_TIMEOUT_SECONDS * 2
+        with self._lock:
+            # Prune workers that have been gone for 2x the timeout
+            stale = [
+                h
+                for h, ts in self._worker_last_seen.items()
+                if (now - ts) > prune_threshold
+            ]
+            for h in stale:
+                del self._worker_last_seen[h]
+
+            result: dict[str, bool] = {}
+            for hostname, last_seen in self._worker_last_seen.items():
+                alive = (now - last_seen) < self._HEARTBEAT_TIMEOUT_SECONDS
+                result[hostname] = alive
+            return result
+
+
+class WorkerHealthCollector(_CachedCollector):
+    """Collects Celery worker health from the heartbeat monitor.
+
+    Reads worker status from ``WorkerHeartbeatMonitor`` which listens
+    to the Celery event stream via a single persistent connection.
+    """
+
+    def __init__(self, cache_ttl: float = 30.0) -> None:
         super().__init__(cache_ttl)
-        self._celery_app: Any | None = None
-        # worker short-name → consecutive miss count.
-        # Workers start at 0 and reset to 0 each time they respond.
-        # Removed after _MAX_CONSECUTIVE_MISSES missed collects.
-        self._known_workers: dict[str, int] = {}
+        self._monitor: WorkerHeartbeatMonitor | None = None
 
-    def set_celery_app(self, app: Any) -> None:
-        """Set the Celery app instance for inspect commands."""
-        self._celery_app = app
+    def set_monitor(self, monitor: WorkerHeartbeatMonitor) -> None:
+        """Set the heartbeat monitor instance."""
+        self._monitor = monitor
 
     def _collect_fresh(self) -> list[GaugeMetricFamily]:
-        if self._celery_app is None:
+        if self._monitor is None:
             return []
 
         active_workers = GaugeMetricFamily(
             "onyx_celery_active_worker_count",
-            "Number of active Celery workers responding to ping",
+            "Number of active Celery workers with recent heartbeats",
         )
         worker_up = GaugeMetricFamily(
             "onyx_celery_worker_up",
@@ -491,37 +578,15 @@ class WorkerHealthCollector(_CachedCollector):
         )
 
         try:
-            inspector = self._celery_app.control.inspect(timeout=3.0)
-            ping_result = inspector.ping()
+            status = self._monitor.get_worker_status()
+            alive_count = sum(1 for alive in status.values() if alive)
+            active_workers.add_metric([], alive_count)
 
-            responding: set[str] = set()
-            if ping_result:
-                active_workers.add_metric([], len(ping_result))
-                for worker_name in ping_result:
-                    # Strip hostname suffix for cleaner labels
-                    short_name = worker_name.split("@")[0]
-                    responding.add(short_name)
-            else:
-                active_workers.add_metric([], 0)
-
-            # Register newly-seen workers and reset miss count for
-            # workers that responded.
-            for short_name in responding:
-                self._known_workers[short_name] = 0
-
-            # Increment miss count for non-responding workers and evict
-            # those that have been missing too long.
-            stale = []
-            for short_name in list(self._known_workers):
-                if short_name not in responding:
-                    self._known_workers[short_name] += 1
-                    if self._known_workers[short_name] >= self._MAX_CONSECUTIVE_MISSES:
-                        stale.append(short_name)
-            for short_name in stale:
-                del self._known_workers[short_name]
-
-            for short_name in sorted(self._known_workers):
-                worker_up.add_metric([short_name], 1 if short_name in responding else 0)
+            for hostname in sorted(status):
+                # Use short name (before @) for single-host deployments,
+                # full hostname when multiple hosts share a worker type.
+                label = hostname.split("@")[0]
+                worker_up.add_metric([label], 1 if status[hostname] else 0)
         except Exception:
             logger.debug("Failed to collect worker health metrics", exc_info=True)
 

backend/onyx/server/metrics/indexing_pipeline_setup.py

@@ -3,24 +3,21 @@
 Called once by the monitoring celery worker after Redis and DB are ready.
 """
 
-from collections.abc import Callable
-from typing import Any
-
 from celery import Celery
 from prometheus_client.registry import REGISTRY
-from redis import Redis
 
 from onyx.server.metrics.indexing_pipeline import ConnectorHealthCollector
 from onyx.server.metrics.indexing_pipeline import IndexAttemptCollector
 from onyx.server.metrics.indexing_pipeline import QueueDepthCollector
 from onyx.server.metrics.indexing_pipeline import RedisHealthCollector
 from onyx.server.metrics.indexing_pipeline import WorkerHealthCollector
+from onyx.server.metrics.indexing_pipeline import WorkerHeartbeatMonitor
 from onyx.utils.logger import setup_logger
 
 logger = setup_logger()
 
 # Module-level singletons — these are lightweight objects (no connections or DB
-# state) until configure() / set_redis_factory() is called. Keeping them at
+# state) until configure() / set_celery_app() is called. Keeping them at
 # module level ensures they survive the lifetime of the worker process and are
 # only registered with the Prometheus registry once.
 _queue_collector = QueueDepthCollector()
@@ -28,75 +25,28 @@ _attempt_collector = IndexAttemptCollector()
 _connector_collector = ConnectorHealthCollector()
 _redis_health_collector = RedisHealthCollector()
 _worker_health_collector = WorkerHealthCollector()
-
-
-def _make_broker_redis_factory(celery_app: Celery) -> Callable[[], Redis]:
-    """Create a factory that returns a cached broker Redis client.
-
-    Reuses a single connection across scrapes to avoid leaking connections.
-    Reconnects automatically if the cached connection becomes stale.
-    """
-    _cached_client: list[Redis | None] = [None]
-    # Keep a reference to the Kombu Connection so we can close it on
-    # reconnect (the raw Redis client outlives the Kombu wrapper).
-    _cached_kombu_conn: list[Any] = [None]
-
-    def _close_client(client: Redis) -> None:
-        """Best-effort close of a Redis client."""
-        try:
-            client.close()
-        except Exception:
-            logger.debug("Failed to close stale Redis client", exc_info=True)
-
-    def _close_kombu_conn() -> None:
-        """Best-effort close of the cached Kombu Connection."""
-        conn = _cached_kombu_conn[0]
-        if conn is not None:
-            try:
-                conn.close()
-            except Exception:
-                logger.debug("Failed to close Kombu connection", exc_info=True)
-            _cached_kombu_conn[0] = None
-
-    def _get_broker_redis() -> Redis:
-        client = _cached_client[0]
-        if client is not None:
-            try:
-                client.ping()
-                return client
-            except Exception:
-                logger.debug("Cached Redis client stale, reconnecting")
-                _close_client(client)
-                _cached_client[0] = None
-                _close_kombu_conn()
-
-        # Get a fresh Redis client from the broker connection.
-        # We hold this client long-term (cached above) rather than using a
-        # context manager, because we need it to persist across scrapes.
-        # The caching logic above ensures we only ever hold one connection,
-        # and we close it explicitly on reconnect.
-        conn = celery_app.broker_connection()
-        # kombu's Channel exposes .client at runtime (the underlying Redis
-        # client) but the type stubs don't declare it.
-        new_client: Redis = conn.channel().client  # type: ignore[attr-defined]
-        _cached_client[0] = new_client
-        _cached_kombu_conn[0] = conn
-        return new_client
-
-    return _get_broker_redis
+_heartbeat_monitor: WorkerHeartbeatMonitor | None = None
 
 
 def setup_indexing_pipeline_metrics(celery_app: Celery) -> None:
     """Register all indexing pipeline collectors with the default registry.
 
     Args:
-        celery_app: The Celery application instance. Used to obtain a fresh
+        celery_app: The Celery application instance. Used to obtain a
             broker Redis client on each scrape for queue depth metrics.
     """
-    redis_factory = _make_broker_redis_factory(celery_app)
-    _queue_collector.set_redis_factory(redis_factory)
-    _redis_health_collector.set_redis_factory(redis_factory)
-    _worker_health_collector.set_celery_app(celery_app)
+    _queue_collector.set_celery_app(celery_app)
+    _redis_health_collector.set_celery_app(celery_app)
+
+    # Start the heartbeat monitor daemon thread — uses a single persistent
+    # connection to receive worker-heartbeat events.
+    # Module-level singleton prevents duplicate threads on re-entry.
+    global _heartbeat_monitor
+    if _heartbeat_monitor is None:
+        _heartbeat_monitor = WorkerHeartbeatMonitor(celery_app)
+        _heartbeat_monitor.start()
+    _worker_health_collector.set_monitor(_heartbeat_monitor)
+
     _attempt_collector.configure()
     _connector_collector.configure()
 

backend/onyx/server/query_and_chat/chat_backend.py

@@ -29,7 +29,6 @@ from onyx.chat.models import ChatFullResponse
 from onyx.chat.models import CreateChatSessionID
 from onyx.chat.process_message import gather_stream_full
 from onyx.chat.process_message import handle_stream_message_objects
-from onyx.chat.process_message import run_multi_model_stream
 from onyx.chat.prompt_utils import get_default_base_system_prompt
 from onyx.chat.stop_signal_checker import set_fence
 from onyx.configs.app_configs import WEB_DOMAIN
@@ -47,7 +46,6 @@ from onyx.db.chat import get_chat_messages_by_session
 from onyx.db.chat import get_chat_session_by_id
 from onyx.db.chat import get_chat_sessions_by_user
 from onyx.db.chat import set_as_latest_chat_message
-from onyx.db.chat import set_preferred_response
 from onyx.db.chat import translate_db_message_to_chat_message_detail
 from onyx.db.chat import update_chat_session
 from onyx.db.chat_search import search_chat_sessions
@@ -62,8 +60,6 @@ from onyx.db.persona import get_persona_by_id
 from onyx.db.usage import increment_usage
 from onyx.db.usage import UsageType
 from onyx.db.user_file import get_file_id_by_user_file_id
-from onyx.error_handling.error_codes import OnyxErrorCode
-from onyx.error_handling.exceptions import OnyxError
 from onyx.file_store.file_store import get_default_file_store
 from onyx.llm.constants import LlmProviderNames
 from onyx.llm.factory import get_default_llm
@@ -85,7 +81,6 @@ from onyx.server.query_and_chat.models import ChatSessionUpdateRequest
 from onyx.server.query_and_chat.models import MessageOrigin
 from onyx.server.query_and_chat.models import RenameChatSessionResponse
 from onyx.server.query_and_chat.models import SendMessageRequest
-from onyx.server.query_and_chat.models import SetPreferredResponseRequest
 from onyx.server.query_and_chat.models import UpdateChatSessionTemperatureRequest
 from onyx.server.query_and_chat.models import UpdateChatSessionThreadRequest
 from onyx.server.query_and_chat.session_loading import (
@@ -575,46 +570,6 @@ def handle_send_chat_message(
     if get_hashed_api_key_from_request(request) or get_hashed_pat_from_request(request):
         chat_message_req.origin = MessageOrigin.API
 
-    # Multi-model streaming path: 2-3 LLMs in parallel (streaming only)
-    is_multi_model = (
-        chat_message_req.llm_overrides is not None
-        and len(chat_message_req.llm_overrides) > 1
-    )
-    if is_multi_model and chat_message_req.stream:
-        # Narrowed here; is_multi_model already checked llm_overrides is not None
-        llm_overrides = chat_message_req.llm_overrides or []
-
-        def multi_model_stream_generator() -> Generator[str, None, None]:
-            try:
-                with get_session_with_current_tenant() as db_session:
-                    for obj in run_multi_model_stream(
-                        new_msg_req=chat_message_req,
-                        user=user,
-                        db_session=db_session,
-                        llm_overrides=llm_overrides,
-                        litellm_additional_headers=extract_headers(
-                            request.headers, LITELLM_PASS_THROUGH_HEADERS
-                        ),
-                        custom_tool_additional_headers=get_custom_tool_additional_request_headers(
-                            request.headers
-                        ),
-                        mcp_headers=chat_message_req.mcp_headers,
-                    ):
-                        yield get_json_line(obj.model_dump())
-            except Exception as e:
-                logger.exception("Error in multi-model streaming")
-                yield json.dumps({"error": str(e)})
-
-        return StreamingResponse(
-            multi_model_stream_generator(), media_type="text/event-stream"
-        )
-
-    if is_multi_model and not chat_message_req.stream:
-        raise OnyxError(
-            OnyxErrorCode.INVALID_INPUT,
-            "Multi-model mode (llm_overrides with >1 entry) requires stream=True.",
-        )
-
     # Non-streaming path: consume all packets and return complete response
     if not chat_message_req.stream:
         with get_session_with_current_tenant() as db_session:
@@ -705,30 +660,6 @@ def set_message_as_latest(
     )
 
 
-@router.put("/set-preferred-response")
-def set_preferred_response_endpoint(
-    request_body: SetPreferredResponseRequest,
-    user: User | None = Depends(current_user),
-    db_session: Session = Depends(get_session),
-) -> None:
-    """Set the preferred assistant response for a multi-model turn."""
-    try:
-        # Ownership check: get_chat_message raises ValueError if the message
-        # doesn't belong to this user, preventing cross-user mutation.
-        get_chat_message(
-            chat_message_id=request_body.user_message_id,
-            user_id=user.id if user else None,
-            db_session=db_session,
-        )
-        set_preferred_response(
-            db_session=db_session,
-            user_message_id=request_body.user_message_id,
-            preferred_assistant_message_id=request_body.preferred_response_id,
-        )
-    except ValueError as e:
-        raise OnyxError(OnyxErrorCode.INVALID_INPUT, str(e))
-
-
 @router.post("/create-chat-message-feedback")
 def create_chat_feedback(
     feedback: ChatFeedbackRequest,

backend/onyx/server/query_and_chat/placement.py

@@ -2,11 +2,25 @@ from pydantic import BaseModel
 
 
 class Placement(BaseModel):
-    # Which iterative block in the UI is this part of, these are ordered and smaller ones happened first
+    """Coordinates that identify where a streaming packet belongs in the UI.
+
+    The frontend uses these fields to route each packet to the correct turn,
+    tool tab, agent sub-turn, and (in multi-model mode) response column.
+
+    Attributes:
+        turn_index: Monotonically increasing index of the iterative reasoning block
+            (e.g. tool call round) within this chat message. Lower values happened first.
+        tab_index: Disambiguates parallel tool calls within the same turn so each
+            tool's output can be displayed in its own tab.
+        sub_turn_index: Nesting level for tools that invoke other tools. ``None`` for
+            top-level packets; an integer for tool-within-tool output.
+        model_index: Which model this packet belongs to. ``0`` for single-model
+            responses; ``0``, ``1``, or ``2`` for multi-model comparison. ``None``
+            for pre-LLM setup packets (e.g. message ID info) that are yielded
+            before any Emitter runs.
+    """
+
     turn_index: int
-    # For parallel tool calls to preserve order of execution
     tab_index: int = 0
-    # Used for tools/agents that call other tools, this currently doesn't support nested agents but can be added later
     sub_turn_index: int | None = None
-    # For multi-model streaming: identifies which model (0, 1, 2) this packet belongs to.
     model_index: int | None = None

backend/onyx/server/settings/api.py

@@ -9,7 +9,9 @@ from onyx import __version__ as onyx_version
 from onyx.auth.users import current_admin_user
 from onyx.auth.users import current_user
 from onyx.auth.users import is_user_admin
+from onyx.configs.app_configs import DEFAULT_USER_FILE_MAX_UPLOAD_SIZE_MB
 from onyx.configs.app_configs import DISABLE_VECTOR_DB
+from onyx.configs.app_configs import MAX_ALLOWED_UPLOAD_SIZE_MB
 from onyx.configs.constants import KV_REINDEX_KEY
 from onyx.configs.constants import NotificationType
 from onyx.db.engine.sql_engine import get_session
@@ -17,10 +19,16 @@ from onyx.db.models import User
 from onyx.db.notification import dismiss_all_notifications
 from onyx.db.notification import get_notifications
 from onyx.db.notification import update_notification_last_shown
+from onyx.error_handling.error_codes import OnyxErrorCode
+from onyx.error_handling.exceptions import OnyxError
 from onyx.hooks.utils import HOOKS_AVAILABLE
 from onyx.key_value_store.factory import get_kv_store
 from onyx.key_value_store.interface import KvKeyNotFoundError
 from onyx.server.features.build.utils import is_onyx_craft_enabled
+from onyx.server.settings.models import (
+    DEFAULT_FILE_TOKEN_COUNT_THRESHOLD_K_NO_VECTOR_DB,
+)
+from onyx.server.settings.models import DEFAULT_FILE_TOKEN_COUNT_THRESHOLD_K_VECTOR_DB
 from onyx.server.settings.models import Notification
 from onyx.server.settings.models import Settings
 from onyx.server.settings.models import UserSettings
@@ -41,6 +49,15 @@ basic_router = APIRouter(prefix="/settings")
 def admin_put_settings(
     settings: Settings, _: User = Depends(current_admin_user)
 ) -> None:
+    if (
+        settings.user_file_max_upload_size_mb is not None
+        and settings.user_file_max_upload_size_mb > 0
+        and settings.user_file_max_upload_size_mb > MAX_ALLOWED_UPLOAD_SIZE_MB
+    ):
+        raise OnyxError(
+            OnyxErrorCode.INVALID_INPUT,
+            f"File upload size limit cannot exceed {MAX_ALLOWED_UPLOAD_SIZE_MB} MB",
+        )
     store_settings(settings)
 
 
@@ -83,6 +100,16 @@ def fetch_settings(
         vector_db_enabled=not DISABLE_VECTOR_DB,
         hooks_enabled=HOOKS_AVAILABLE,
         version=onyx_version,
+        max_allowed_upload_size_mb=MAX_ALLOWED_UPLOAD_SIZE_MB,
+        default_user_file_max_upload_size_mb=min(
+            DEFAULT_USER_FILE_MAX_UPLOAD_SIZE_MB,
+            MAX_ALLOWED_UPLOAD_SIZE_MB,
+        ),
+        default_file_token_count_threshold_k=(
+            DEFAULT_FILE_TOKEN_COUNT_THRESHOLD_K_NO_VECTOR_DB
+            if DISABLE_VECTOR_DB
+            else DEFAULT_FILE_TOKEN_COUNT_THRESHOLD_K_VECTOR_DB
+        ),
     )
 
 

backend/onyx/server/settings/models.py

@@ -2,12 +2,19 @@ from datetime import datetime
 from enum import Enum
 
 from pydantic import BaseModel
+from pydantic import Field
 
+from onyx.configs.app_configs import DEFAULT_USER_FILE_MAX_UPLOAD_SIZE_MB
+from onyx.configs.app_configs import DISABLE_VECTOR_DB
+from onyx.configs.app_configs import MAX_ALLOWED_UPLOAD_SIZE_MB
 from onyx.configs.constants import NotificationType
 from onyx.configs.constants import QueryHistoryType
 from onyx.db.models import Notification as NotificationDBModel
 from shared_configs.configs import POSTGRES_DEFAULT_SCHEMA
 
+DEFAULT_FILE_TOKEN_COUNT_THRESHOLD_K_VECTOR_DB = 200
+DEFAULT_FILE_TOKEN_COUNT_THRESHOLD_K_NO_VECTOR_DB = 10000
+
 
 class PageType(str, Enum):
     CHAT = "chat"
@@ -78,7 +85,12 @@ class Settings(BaseModel):
 
     # User Knowledge settings
     user_knowledge_enabled: bool | None = True
-    user_file_max_upload_size_mb: int | None = None
+    user_file_max_upload_size_mb: int | None = Field(
+        default=DEFAULT_USER_FILE_MAX_UPLOAD_SIZE_MB, ge=0
+    )
+    file_token_count_threshold_k: int | None = Field(
+        default=None, ge=0  # thousands of tokens; None = context-aware default
+    )
 
     # Connector settings
     show_extra_connectors: bool | None = True
@@ -108,3 +120,14 @@ class UserSettings(Settings):
     hooks_enabled: bool = False
     # Application version, read from the ONYX_VERSION env var at startup.
     version: str | None = None
+    # Hard ceiling for user_file_max_upload_size_mb, derived from env var.
+    max_allowed_upload_size_mb: int = MAX_ALLOWED_UPLOAD_SIZE_MB
+    # Factory defaults so the frontend can show a "restore default" button.
+    default_user_file_max_upload_size_mb: int = DEFAULT_USER_FILE_MAX_UPLOAD_SIZE_MB
+    default_file_token_count_threshold_k: int = Field(
+        default_factory=lambda: (
+            DEFAULT_FILE_TOKEN_COUNT_THRESHOLD_K_NO_VECTOR_DB
+            if DISABLE_VECTOR_DB
+            else DEFAULT_FILE_TOKEN_COUNT_THRESHOLD_K_VECTOR_DB
+        )
+    )

backend/onyx/server/settings/store.py

@@ -1,13 +1,19 @@
 from onyx.cache.factory import get_cache_backend
+from onyx.configs.app_configs import DEFAULT_USER_FILE_MAX_UPLOAD_SIZE_MB
 from onyx.configs.app_configs import DISABLE_USER_KNOWLEDGE
+from onyx.configs.app_configs import DISABLE_VECTOR_DB
 from onyx.configs.app_configs import ENABLE_OPENSEARCH_INDEXING_FOR_ONYX
+from onyx.configs.app_configs import MAX_ALLOWED_UPLOAD_SIZE_MB
 from onyx.configs.app_configs import ONYX_QUERY_HISTORY_TYPE
 from onyx.configs.app_configs import SHOW_EXTRA_CONNECTORS
-from onyx.configs.app_configs import USER_FILE_MAX_UPLOAD_SIZE_MB
 from onyx.configs.constants import KV_SETTINGS_KEY
 from onyx.configs.constants import OnyxRedisLocks
 from onyx.key_value_store.factory import get_kv_store
 from onyx.key_value_store.interface import KvKeyNotFoundError
+from onyx.server.settings.models import (
+    DEFAULT_FILE_TOKEN_COUNT_THRESHOLD_K_NO_VECTOR_DB,
+)
+from onyx.server.settings.models import DEFAULT_FILE_TOKEN_COUNT_THRESHOLD_K_VECTOR_DB
 from onyx.server.settings.models import Settings
 from onyx.utils.logger import setup_logger
 
@@ -51,9 +57,36 @@ def load_settings() -> Settings:
     if DISABLE_USER_KNOWLEDGE:
         settings.user_knowledge_enabled = False
 
-    settings.user_file_max_upload_size_mb = USER_FILE_MAX_UPLOAD_SIZE_MB
     settings.show_extra_connectors = SHOW_EXTRA_CONNECTORS
     settings.opensearch_indexing_enabled = ENABLE_OPENSEARCH_INDEXING_FOR_ONYX
+
+    # Resolve context-aware defaults for token threshold.
+    # None = admin hasn't set a value yet → use context-aware default.
+    # 0 = admin explicitly chose "no limit" → preserve as-is.
+    if settings.file_token_count_threshold_k is None:
+        settings.file_token_count_threshold_k = (
+            DEFAULT_FILE_TOKEN_COUNT_THRESHOLD_K_NO_VECTOR_DB
+            if DISABLE_VECTOR_DB
+            else DEFAULT_FILE_TOKEN_COUNT_THRESHOLD_K_VECTOR_DB
+        )
+
+    # Upload size: 0 and None are treated as "unset" (not "no limit") →
+    # fall back to min(configured default, hard ceiling).
+    if not settings.user_file_max_upload_size_mb:
+        settings.user_file_max_upload_size_mb = min(
+            DEFAULT_USER_FILE_MAX_UPLOAD_SIZE_MB,
+            MAX_ALLOWED_UPLOAD_SIZE_MB,
+        )
+
+    # Clamp to env ceiling so stale KV values are capped even if the
+    # operator lowered MAX_ALLOWED_UPLOAD_SIZE_MB after a higher value
+    # was already saved (api.py only guards new writes).
+    if (
+        settings.user_file_max_upload_size_mb > 0
+        and settings.user_file_max_upload_size_mb > MAX_ALLOWED_UPLOAD_SIZE_MB
+    ):
+        settings.user_file_max_upload_size_mb = MAX_ALLOWED_UPLOAD_SIZE_MB
+
     return settings
 
 

backend/onyx/tools/fake_tools/research_agent.py

@@ -1,3 +1,4 @@
+import queue
 import time
 from collections.abc import Callable
 from typing import Any
@@ -708,7 +709,6 @@ def run_research_agent_calls(
 
 
 if __name__ == "__main__":
-    from queue import Queue
     from uuid import uuid4
 
     from onyx.chat.chat_state import ChatStateContainer
@@ -744,8 +744,8 @@ if __name__ == "__main__":
         if user is None:
             raise ValueError("No users found in database. Please create a user first.")
 
-        bus: Queue[Packet] = Queue()
-        emitter = Emitter(bus)
+        emitter_queue: queue.Queue = queue.Queue()
+        emitter = Emitter(merged_queue=emitter_queue)
         state_container = ChatStateContainer()
 
         tool_dict = construct_tools(
@@ -792,4 +792,4 @@ if __name__ == "__main__":
             print(result.intermediate_report)
             print("=" * 80)
             print(f"Citations: {result.citation_mapping}")
-            print(f"Total packets emitted: {bus.qsize()}")
+            print(f"Total packets emitted: {emitter_queue.qsize()}")

backend/onyx/tools/tool_implementations/custom/custom_tool.py

@@ -1,5 +1,6 @@
 import csv
 import json
+import queue
 import uuid
 from io import BytesIO
 from io import StringIO
@@ -11,7 +12,6 @@ import requests
 from requests import JSONDecodeError
 
 from onyx.chat.emitter import Emitter
-from onyx.chat.emitter import get_default_emitter
 from onyx.configs.constants import FileOrigin
 from onyx.file_store.file_store import get_default_file_store
 from onyx.server.query_and_chat.placement import Placement
@@ -296,9 +296,9 @@ def build_custom_tools_from_openapi_schema_and_headers(
     url = openapi_to_url(openapi_schema)
     method_specs = openapi_to_method_specs(openapi_schema)
 
-    # Use default emitter if none provided
+    # Use a discard emitter if none provided (packets go nowhere)
     if emitter is None:
-        emitter = get_default_emitter()
+        emitter = Emitter(merged_queue=queue.Queue())
 
     return [
         CustomTool(
@@ -367,7 +367,7 @@ if __name__ == "__main__":
     tools = build_custom_tools_from_openapi_schema_and_headers(
         tool_id=0,  # dummy tool id
         openapi_schema=openapi_schema,
-        emitter=get_default_emitter(),
+        emitter=Emitter(merged_queue=queue.Queue()),
         dynamic_schema_info=None,
     )
 

backend/requirements/default.txt

@@ -187,7 +187,7 @@ coloredlogs==15.0.1
     # via onnxruntime
 courlan==1.3.2
     # via trafilatura
-cryptography==46.0.5
+cryptography==46.0.6
     # via
     #   authlib
     #   google-auth
@@ -449,7 +449,7 @@ kombu==5.5.4
     # via celery
 kubernetes==31.0.0
     # via onyx
-langchain-core==1.2.11
+langchain-core==1.2.22
     # via onyx
 langdetect==1.0.9
     # via unstructured
@@ -735,7 +735,7 @@ pyee==13.0.0
     # via playwright
 pygithub==2.5.0
     # via onyx
-pygments==2.19.2
+pygments==2.20.0
     # via rich
 pyjwt==2.12.0
     # via

backend/requirements/dev.txt

@@ -97,7 +97,7 @@ comm==0.2.3
     # via ipykernel
 contourpy==1.3.3
     # via matplotlib
-cryptography==46.0.5
+cryptography==46.0.6
     # via
     #   google-auth
     #   pyjwt
@@ -263,7 +263,7 @@ oauthlib==3.2.2
     # via
     #   kubernetes
     #   requests-oauthlib
-onyx-devtools==0.7.1
+onyx-devtools==0.7.2
     # via onyx
 openai==2.14.0
     # via
@@ -349,7 +349,7 @@ pydantic-core==2.33.2
     # via pydantic
 pydantic-settings==2.12.0
     # via mcp
-pygments==2.19.2
+pygments==2.20.0
     # via
     #   ipython
     #   ipython-pygments-lexers

backend/requirements/ee.txt

@@ -76,7 +76,7 @@ colorama==0.4.6 ; sys_platform == 'win32'
     # via
     #   click
     #   tqdm
-cryptography==46.0.5
+cryptography==46.0.6
     # via
     #   google-auth
     #   pyjwt

backend/requirements/model_server.txt

@@ -92,7 +92,7 @@ colorama==0.4.6 ; sys_platform == 'win32'
     # via
     #   click
     #   tqdm
-cryptography==46.0.5
+cryptography==46.0.6
     # via
     #   google-auth
     #   pyjwt

backend/shared_configs/configs.py

@@ -191,25 +191,6 @@ IGNORED_SYNCING_TENANT_LIST = (
     else None
 )
 
-# Global flag to skip userfile threshold for all users/tenants
-SKIP_USERFILE_THRESHOLD = (
-    os.environ.get("SKIP_USERFILE_THRESHOLD", "").lower() == "true"
-)
-
-# Comma-separated list of specific tenant IDs to skip threshold (multi-tenant only)
-SKIP_USERFILE_THRESHOLD_TENANT_IDS = os.environ.get(
-    "SKIP_USERFILE_THRESHOLD_TENANT_IDS"
-)
-SKIP_USERFILE_THRESHOLD_TENANT_LIST = (
-    [
-        tenant.strip()
-        for tenant in SKIP_USERFILE_THRESHOLD_TENANT_IDS.split(",")
-        if tenant.strip()
-    ]
-    if SKIP_USERFILE_THRESHOLD_TENANT_IDS
-    else None
-)
-
 ENVIRONMENT = os.environ.get("ENVIRONMENT") or "not_explicitly_set"
 
 

backend/tests/daily/connectors/slack/test_slack_perm_sync.py

@@ -1,4 +1,6 @@
 import time
+from datetime import datetime
+from datetime import timezone
 
 import pytest
 
@@ -17,6 +19,10 @@ PRIVATE_CHANNEL_USERS = [
     "test_user_2@onyx-test.com",
 ]
 
+# Predates any test workspace messages, so the result set should match
+# the "no start time" case while exercising the oldest= parameter.
+OLDEST_TS_2016 = datetime(2016, 1, 1, tzinfo=timezone.utc).timestamp()
+
 pytestmark = pytest.mark.usefixtures("enable_ee")
 
 
@@ -105,15 +111,17 @@ def test_load_from_checkpoint_access__private_channel(
     ],
     indirect=True,
 )
+@pytest.mark.parametrize("start_ts", [None, OLDEST_TS_2016])
 def test_slim_documents_access__public_channel(
     slack_connector: SlackConnector,
+    start_ts: float | None,
 ) -> None:
     """Test that retrieve_all_slim_docs_perm_sync returns correct access information for slim documents."""
     if not slack_connector.client:
         raise RuntimeError("Web client must be defined")
 
     slim_docs_generator = slack_connector.retrieve_all_slim_docs_perm_sync(
-        start=0.0,
+        start=start_ts,
         end=time.time(),
     )
 
@@ -149,7 +157,7 @@ def test_slim_documents_access__private_channel(
         raise RuntimeError("Web client must be defined")
 
     slim_docs_generator = slack_connector.retrieve_all_slim_docs_perm_sync(
-        start=0.0,
+        start=None,
         end=time.time(),
     )
 

backend/tests/external_dependency_unit/answer/stream_test_utils.py

@@ -27,11 +27,13 @@ def create_placement(
     turn_index: int,
     tab_index: int = 0,
     sub_turn_index: int | None = None,
+    model_index: int | None = 0,
 ) -> Placement:
     return Placement(
         turn_index=turn_index,
         tab_index=tab_index,
         sub_turn_index=sub_turn_index,
+        model_index=model_index,
     )
 
 

backend/tests/external_dependency_unit/celery/test_persona_file_sync.py

@@ -129,6 +129,10 @@ def _patch_task_app(task: Any, mock_app: MagicMock) -> Generator[None, None, Non
             return_value=mock_app,
         ),
         patch(_PATCH_QUEUE_DEPTH, return_value=0),
+        patch(
+            "onyx.background.celery.tasks.user_file_processing.tasks.celery_get_broker_client",
+            return_value=MagicMock(),
+        ),
     ):
         yield
 

backend/tests/external_dependency_unit/celery/test_user_file_delete_queue.py

@@ -88,10 +88,22 @@ def _patch_task_app(task: Any, mock_app: MagicMock) -> Generator[None, None, Non
     the actual task instance.  We patch ``app`` on that instance's class
     (a unique Celery-generated Task subclass) so the mock is scoped to this
     task only.
+
+    Also patches ``celery_get_broker_client`` so the mock app doesn't need
+    a real broker URL.
     """
     task_instance = task.run.__self__
-    with patch.object(
-        type(task_instance), "app", new_callable=PropertyMock, return_value=mock_app
+    with (
+        patch.object(
+            type(task_instance),
+            "app",
+            new_callable=PropertyMock,
+            return_value=mock_app,
+        ),
+        patch(
+            "onyx.background.celery.tasks.user_file_processing.tasks.celery_get_broker_client",
+            return_value=MagicMock(),
+        ),
     ):
         yield
 

backend/tests/external_dependency_unit/celery/test_user_file_indexing_adapter.py

@@ -1,7 +1,7 @@
 """
 External dependency unit tests for UserFileIndexingAdapter metadata writing.
 
-Validates that build_metadata_aware_chunks produces DocMetadataAwareIndexChunk
+Validates that prepare_enrichment produces DocMetadataAwareIndexChunk
 objects with both `user_project` and `personas` fields populated correctly
 based on actual DB associations.
 
@@ -127,7 +127,7 @@ def _make_index_chunk(user_file: UserFile) -> IndexChunk:
 
 
 class TestAdapterWritesBothMetadataFields:
-    """build_metadata_aware_chunks must populate user_project AND personas."""
+    """prepare_enrichment must populate user_project AND personas."""
 
     @patch(
         "onyx.indexing.adapters.user_file_indexing_adapter.get_default_llm",
@@ -153,15 +153,13 @@ class TestAdapterWritesBothMetadataFields:
         doc = chunk.source_document
         context = DocumentBatchPrepareContext(updatable_docs=[doc], id_to_boost_map={})
 
-        result = adapter.build_metadata_aware_chunks(
-            chunks_with_embeddings=[chunk],
-            chunk_content_scores=[1.0],
-            tenant_id=TEST_TENANT_ID,
+        enricher = adapter.prepare_enrichment(
             context=context,
+            tenant_id=TEST_TENANT_ID,
+            chunks=[chunk],
         )
+        aware_chunk = enricher.enrich_chunk(chunk, 1.0)
 
-        assert len(result.chunks) == 1
-        aware_chunk = result.chunks[0]
         assert persona.id in aware_chunk.personas
         assert aware_chunk.user_project == []
 
@@ -190,15 +188,13 @@ class TestAdapterWritesBothMetadataFields:
             updatable_docs=[chunk.source_document], id_to_boost_map={}
         )
 
-        result = adapter.build_metadata_aware_chunks(
-            chunks_with_embeddings=[chunk],
-            chunk_content_scores=[1.0],
-            tenant_id=TEST_TENANT_ID,
+        enricher = adapter.prepare_enrichment(
             context=context,
+            tenant_id=TEST_TENANT_ID,
+            chunks=[chunk],
         )
+        aware_chunk = enricher.enrich_chunk(chunk, 1.0)
 
-        assert len(result.chunks) == 1
-        aware_chunk = result.chunks[0]
         assert project.id in aware_chunk.user_project
         assert aware_chunk.personas == []
 
@@ -229,14 +225,13 @@ class TestAdapterWritesBothMetadataFields:
             updatable_docs=[chunk.source_document], id_to_boost_map={}
         )
 
-        result = adapter.build_metadata_aware_chunks(
-            chunks_with_embeddings=[chunk],
-            chunk_content_scores=[1.0],
-            tenant_id=TEST_TENANT_ID,
+        enricher = adapter.prepare_enrichment(
             context=context,
+            tenant_id=TEST_TENANT_ID,
+            chunks=[chunk],
         )
+        aware_chunk = enricher.enrich_chunk(chunk, 1.0)
 
-        aware_chunk = result.chunks[0]
         assert persona.id in aware_chunk.personas
         assert project.id in aware_chunk.user_project
 
@@ -261,14 +256,13 @@ class TestAdapterWritesBothMetadataFields:
             updatable_docs=[chunk.source_document], id_to_boost_map={}
         )
 
-        result = adapter.build_metadata_aware_chunks(
-            chunks_with_embeddings=[chunk],
-            chunk_content_scores=[1.0],
-            tenant_id=TEST_TENANT_ID,
+        enricher = adapter.prepare_enrichment(
             context=context,
+            tenant_id=TEST_TENANT_ID,
+            chunks=[chunk],
         )
+        aware_chunk = enricher.enrich_chunk(chunk, 1.0)
 
-        aware_chunk = result.chunks[0]
         assert aware_chunk.personas == []
         assert aware_chunk.user_project == []
 
@@ -300,12 +294,11 @@ class TestAdapterWritesBothMetadataFields:
             updatable_docs=[chunk.source_document], id_to_boost_map={}
         )
 
-        result = adapter.build_metadata_aware_chunks(
-            chunks_with_embeddings=[chunk],
-            chunk_content_scores=[1.0],
-            tenant_id=TEST_TENANT_ID,
+        enricher = adapter.prepare_enrichment(
             context=context,
+            tenant_id=TEST_TENANT_ID,
+            chunks=[chunk],
         )
+        aware_chunk = enricher.enrich_chunk(chunk, 1.0)
 
-        aware_chunk = result.chunks[0]
         assert set(aware_chunk.personas) == {persona_a.id, persona_b.id}

backend/tests/external_dependency_unit/celery/test_user_file_processing_queue.py

@@ -90,8 +90,17 @@ def _patch_task_app(task: Any, mock_app: MagicMock) -> Generator[None, None, Non
     task only.
     """
     task_instance = task.run.__self__
-    with patch.object(
-        type(task_instance), "app", new_callable=PropertyMock, return_value=mock_app
+    with (
+        patch.object(
+            type(task_instance),
+            "app",
+            new_callable=PropertyMock,
+            return_value=mock_app,
+        ),
+        patch(
+            "onyx.background.celery.tasks.user_file_processing.tasks.celery_get_broker_client",
+            return_value=MagicMock(),
+        ),
     ):
         yield
 

backend/tests/external_dependency_unit/connectors/confluence/test_confluence_group_sync.py

@@ -103,6 +103,11 @@ _EXPECTED_CONFLUENCE_GROUPS = [
         user_emails={"oauth@onyx.app"},
         gives_anyone_access=False,
     ),
+    ExternalUserGroupSet(
+        id="no yuhong allowed",
+        user_emails={"hagen@danswer.ai", "pablo@onyx.app", "chris@onyx.app"},
+        gives_anyone_access=False,
+    ),
 ]
 
 

backend/tests/external_dependency_unit/document_index/test_document_index.py

@@ -6,6 +6,7 @@ These tests assume Vespa and OpenSearch are running.
 import time
 import uuid
 from collections.abc import Generator
+from collections.abc import Iterator
 
 import httpx
 import pytest
@@ -21,6 +22,7 @@ from onyx.document_index.opensearch.opensearch_document_index import (
 )
 from onyx.document_index.vespa.index import VespaIndex
 from onyx.document_index.vespa.vespa_document_index import VespaDocumentIndex
+from onyx.indexing.models import DocMetadataAwareIndexChunk
 from tests.external_dependency_unit.constants import TEST_TENANT_ID
 from tests.external_dependency_unit.document_index.conftest import EMBEDDING_DIM
 from tests.external_dependency_unit.document_index.conftest import make_chunk
@@ -201,3 +203,25 @@ class TestDocumentIndexNew:
             assert len(result_map) == 2
             assert result_map[existing_doc] is True
             assert result_map[new_doc] is False
+
+    def test_index_accepts_generator(
+        self,
+        document_indices: list[DocumentIndexNew],
+        tenant_context: None,  # noqa: ARG002
+    ) -> None:
+        """index() accepts a generator (any iterable), not just a list."""
+        for document_index in document_indices:
+            doc_id = f"test_gen_{uuid.uuid4().hex[:8]}"
+            metadata = make_indexing_metadata([doc_id], old_counts=[0], new_counts=[3])
+
+            def chunk_gen() -> Iterator[DocMetadataAwareIndexChunk]:
+                for i in range(3):
+                    yield make_chunk(doc_id, chunk_id=i)
+
+            results = document_index.index(
+                chunks=chunk_gen(), indexing_metadata=metadata
+            )
+
+            assert len(results) == 1
+            assert results[0].document_id == doc_id
+            assert results[0].already_existed is False

backend/tests/external_dependency_unit/document_index/test_document_index_old.py

@@ -5,6 +5,7 @@ These tests assume Vespa and OpenSearch are running.
 
 import time
 from collections.abc import Generator
+from collections.abc import Iterator
 
 import pytest
 
@@ -166,3 +167,29 @@ class TestDocumentIndexOld:
                 batch_retrieval=True,
             )
             assert len(inference_chunks) == 0
+
+    def test_index_accepts_generator(
+        self,
+        document_indices: list[DocumentIndex],
+        tenant_context: None,  # noqa: ARG002
+    ) -> None:
+        """index() accepts a generator (any iterable), not just a list."""
+        for document_index in document_indices:
+
+            def chunk_gen() -> Iterator[DocMetadataAwareIndexChunk]:
+                for i in range(3):
+                    yield make_chunk("test_doc_gen", chunk_id=i)
+
+            index_batch_params = IndexBatchParams(
+                doc_id_to_previous_chunk_cnt={"test_doc_gen": 0},
+                doc_id_to_new_chunk_cnt={"test_doc_gen": 3},
+                tenant_id=get_current_tenant_id(),
+                large_chunks_enabled=False,
+            )
+
+            results = document_index.index(chunk_gen(), index_batch_params)
+
+            assert len(results) == 1
+            record = results.pop()
+            assert record.document_id == "test_doc_gen"
+            assert record.already_existed is False

backend/tests/external_dependency_unit/tools/test_mcp_passthrough_oauth.py

@@ -13,6 +13,7 @@ This test:
 All external HTTP calls are mocked, but Postgres and Redis are running.
 """
 
+import queue
 from typing import Any
 from unittest.mock import patch
 from uuid import uuid4
@@ -20,7 +21,7 @@ from uuid import uuid4
 import pytest
 from sqlalchemy.orm import Session
 
-from onyx.chat.emitter import get_default_emitter
+from onyx.chat.emitter import Emitter
 from onyx.db.enums import MCPAuthenticationPerformer
 from onyx.db.enums import MCPAuthenticationType
 from onyx.db.enums import MCPTransport
@@ -137,7 +138,7 @@ class TestMCPPassThroughOAuth:
         tool_dict = construct_tools(
             persona=persona,
             db_session=db_session,
-            emitter=get_default_emitter(),
+            emitter=Emitter(merged_queue=queue.Queue()),
             user=user,
             llm=llm,
             search_tool_config=search_tool_config,
@@ -200,7 +201,7 @@ class TestMCPPassThroughOAuth:
         tool_dict = construct_tools(
             persona=persona,
             db_session=db_session,
-            emitter=get_default_emitter(),
+            emitter=Emitter(merged_queue=queue.Queue()),
             user=user,
             llm=llm,
             search_tool_config=SearchToolConfig(),
@@ -275,7 +276,7 @@ class TestMCPPassThroughOAuth:
         tool_dict = construct_tools(
             persona=persona,
             db_session=db_session,
-            emitter=get_default_emitter(),
+            emitter=Emitter(merged_queue=queue.Queue()),
             user=user,
             llm=llm,
             search_tool_config=SearchToolConfig(),
@@ -350,7 +351,7 @@ class TestMCPPassThroughOAuth:
         tool_dict = construct_tools(
             persona=persona,
             db_session=db_session,
-            emitter=get_default_emitter(),
+            emitter=Emitter(merged_queue=queue.Queue()),
             user=user,
             llm=llm,
             search_tool_config=SearchToolConfig(),
@@ -458,7 +459,7 @@ class TestMCPPassThroughOAuth:
         tool_dict = construct_tools(
             persona=persona,
             db_session=db_session,
-            emitter=get_default_emitter(),
+            emitter=Emitter(merged_queue=queue.Queue()),
             user=user,
             llm=llm,
             search_tool_config=SearchToolConfig(),
@@ -541,7 +542,7 @@ class TestMCPPassThroughOAuth:
         tool_dict = construct_tools(
             persona=persona,
             db_session=db_session,
-            emitter=get_default_emitter(),
+            emitter=Emitter(merged_queue=queue.Queue()),
             user=user,
             llm=llm,
             search_tool_config=SearchToolConfig(),

backend/tests/external_dependency_unit/tools/test_oauth_tool_integration.py

@@ -8,6 +8,7 @@ Tests the priority logic for OAuth tokens when constructing custom tools:
 All external HTTP calls are mocked, but Postgres and Redis are running.
 """
 
+import queue
 from typing import Any
 from unittest.mock import Mock
 from unittest.mock import patch
@@ -16,7 +17,7 @@ from uuid import uuid4
 import pytest
 from sqlalchemy.orm import Session
 
-from onyx.chat.emitter import get_default_emitter
+from onyx.chat.emitter import Emitter
 from onyx.db.models import OAuthAccount
 from onyx.db.models import OAuthConfig
 from onyx.db.models import Persona
@@ -174,7 +175,7 @@ class TestOAuthToolIntegrationPriority:
         tool_dict = construct_tools(
             persona=persona,
             db_session=db_session,
-            emitter=get_default_emitter(),
+            emitter=Emitter(merged_queue=queue.Queue()),
             user=user,
             llm=llm,
             search_tool_config=search_tool_config,
@@ -232,7 +233,7 @@ class TestOAuthToolIntegrationPriority:
         tool_dict = construct_tools(
             persona=persona,
             db_session=db_session,
-            emitter=get_default_emitter(),
+            emitter=Emitter(merged_queue=queue.Queue()),
             user=user,
             llm=llm,
         )
@@ -284,7 +285,7 @@ class TestOAuthToolIntegrationPriority:
             tool_dict = construct_tools(
                 persona=persona,
                 db_session=db_session,
-                emitter=get_default_emitter(),
+                emitter=Emitter(merged_queue=queue.Queue()),
                 user=user,
                 llm=llm,
             )
@@ -345,7 +346,7 @@ class TestOAuthToolIntegrationPriority:
         tool_dict = construct_tools(
             persona=persona,
             db_session=db_session,
-            emitter=get_default_emitter(),
+            emitter=Emitter(merged_queue=queue.Queue()),
             user=user,
             llm=llm,
         )
@@ -416,7 +417,7 @@ class TestOAuthToolIntegrationPriority:
             tool_dict = construct_tools(
                 persona=persona,
                 db_session=db_session,
-                emitter=get_default_emitter(),
+                emitter=Emitter(merged_queue=queue.Queue()),
                 user=user,
                 llm=llm,
             )
@@ -483,7 +484,7 @@ class TestOAuthToolIntegrationPriority:
         tool_dict = construct_tools(
             persona=persona,
             db_session=db_session,
-            emitter=get_default_emitter(),
+            emitter=Emitter(merged_queue=queue.Queue()),
             user=user,
             llm=llm,
         )
@@ -536,7 +537,7 @@ class TestOAuthToolIntegrationPriority:
         tool_dict = construct_tools(
             persona=persona,
             db_session=db_session,
-            emitter=get_default_emitter(),
+            emitter=Emitter(merged_queue=queue.Queue()),
             user=user,
             llm=llm,
         )

backend/tests/unit/onyx/background/celery/test_celery_redis.py

@@ -0,0 +1,87 @@
+"""Tests for celery_get_broker_client singleton."""
+
+from collections.abc import Iterator
+from unittest.mock import MagicMock
+from unittest.mock import patch
+
+import pytest
+
+from onyx.background.celery import celery_redis
+
+
+@pytest.fixture(autouse=True)
+def reset_singleton() -> Iterator[None]:
+    """Reset the module-level singleton between tests."""
+    celery_redis._broker_client = None
+    celery_redis._broker_url = None
+    yield
+    celery_redis._broker_client = None
+    celery_redis._broker_url = None
+
+
+def _make_mock_app(broker_url: str = "redis://localhost:6379/15") -> MagicMock:
+    app = MagicMock()
+    app.conf.broker_url = broker_url
+    return app
+
+
+class TestCeleryGetBrokerClient:
+    @patch("onyx.background.celery.celery_redis.Redis")
+    def test_creates_client_on_first_call(self, mock_redis_cls: MagicMock) -> None:
+        mock_client = MagicMock()
+        mock_redis_cls.from_url.return_value = mock_client
+
+        app = _make_mock_app()
+        result = celery_redis.celery_get_broker_client(app)
+
+        assert result is mock_client
+        call_args = mock_redis_cls.from_url.call_args
+        assert call_args[0][0] == "redis://localhost:6379/15"
+        assert call_args[1]["decode_responses"] is False
+        assert call_args[1]["socket_keepalive"] is True
+        assert call_args[1]["retry_on_timeout"] is True
+
+    @patch("onyx.background.celery.celery_redis.Redis")
+    def test_reuses_cached_client(self, mock_redis_cls: MagicMock) -> None:
+        mock_client = MagicMock()
+        mock_client.ping.return_value = True
+        mock_redis_cls.from_url.return_value = mock_client
+
+        app = _make_mock_app()
+        client1 = celery_redis.celery_get_broker_client(app)
+        client2 = celery_redis.celery_get_broker_client(app)
+
+        assert client1 is client2
+        # from_url called only once
+        assert mock_redis_cls.from_url.call_count == 1
+
+    @patch("onyx.background.celery.celery_redis.Redis")
+    def test_reconnects_on_ping_failure(self, mock_redis_cls: MagicMock) -> None:
+        stale_client = MagicMock()
+        stale_client.ping.side_effect = ConnectionError("disconnected")
+
+        fresh_client = MagicMock()
+        fresh_client.ping.return_value = True
+
+        mock_redis_cls.from_url.side_effect = [stale_client, fresh_client]
+
+        app = _make_mock_app()
+
+        # First call creates stale_client
+        client1 = celery_redis.celery_get_broker_client(app)
+        assert client1 is stale_client
+
+        # Second call: ping fails, creates fresh_client
+        client2 = celery_redis.celery_get_broker_client(app)
+        assert client2 is fresh_client
+        assert mock_redis_cls.from_url.call_count == 2
+
+    @patch("onyx.background.celery.celery_redis.Redis")
+    def test_uses_broker_url_from_app_config(self, mock_redis_cls: MagicMock) -> None:
+        mock_redis_cls.from_url.return_value = MagicMock()
+
+        app = _make_mock_app("redis://custom-host:6380/3")
+        celery_redis.celery_get_broker_client(app)
+
+        call_args = mock_redis_cls.from_url.call_args
+        assert call_args[0][0] == "redis://custom-host:6380/3"

backend/tests/unit/onyx/chat/test_emitter.py

@@ -0,0 +1,173 @@
+"""Unit tests for the Emitter class.
+
+All tests use the streaming mode (merged_queue required). Emitter has a single
+code path — no standalone bus.
+"""
+
+import queue
+
+from onyx.chat.emitter import Emitter
+from onyx.server.query_and_chat.placement import Placement
+from onyx.server.query_and_chat.streaming_models import OverallStop
+from onyx.server.query_and_chat.streaming_models import Packet
+from onyx.server.query_and_chat.streaming_models import ReasoningStart
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+
+def _placement(
+    turn_index: int = 0,
+    tab_index: int = 0,
+    sub_turn_index: int | None = None,
+) -> Placement:
+    return Placement(
+        turn_index=turn_index,
+        tab_index=tab_index,
+        sub_turn_index=sub_turn_index,
+    )
+
+
+def _packet(
+    turn_index: int = 0,
+    tab_index: int = 0,
+    sub_turn_index: int | None = None,
+) -> Packet:
+    """Build a minimal valid packet with an OverallStop payload."""
+    return Packet(
+        placement=_placement(turn_index, tab_index, sub_turn_index),
+        obj=OverallStop(stop_reason="test"),
+    )
+
+
+def _make_emitter(model_idx: int = 0) -> tuple["Emitter", "queue.Queue"]:
+    """Return (emitter, queue) wired together."""
+    mq: queue.Queue = queue.Queue()
+    return Emitter(merged_queue=mq, model_idx=model_idx), mq
+
+
+# ---------------------------------------------------------------------------
+# Queue routing
+# ---------------------------------------------------------------------------
+
+
+class TestEmitterQueueRouting:
+    def test_emit_lands_on_merged_queue(self) -> None:
+        emitter, mq = _make_emitter()
+        emitter.emit(_packet())
+        assert not mq.empty()
+
+    def test_queue_item_is_tuple_of_key_and_packet(self) -> None:
+        emitter, mq = _make_emitter(model_idx=1)
+        emitter.emit(_packet())
+        item = mq.get_nowait()
+        assert isinstance(item, tuple)
+        assert len(item) == 2
+
+    def test_multiple_packets_delivered_fifo(self) -> None:
+        emitter, mq = _make_emitter()
+        p1 = _packet(turn_index=0)
+        p2 = _packet(turn_index=1)
+        emitter.emit(p1)
+        emitter.emit(p2)
+        _, t1 = mq.get_nowait()
+        _, t2 = mq.get_nowait()
+        assert t1.placement.turn_index == 0
+        assert t2.placement.turn_index == 1
+
+
+# ---------------------------------------------------------------------------
+# model_index tagging
+# ---------------------------------------------------------------------------
+
+
+class TestEmitterModelIndexTagging:
+    def test_n1_default_model_idx_tags_model_index_zero(self) -> None:
+        """N=1: default model_idx=0, so packet gets model_index=0."""
+        emitter, mq = _make_emitter(model_idx=0)
+        emitter.emit(_packet())
+        _key, tagged = mq.get_nowait()
+        assert tagged.placement.model_index == 0
+
+    def test_model_idx_one_tags_packet(self) -> None:
+        emitter, mq = _make_emitter(model_idx=1)
+        emitter.emit(_packet())
+        _key, tagged = mq.get_nowait()
+        assert tagged.placement.model_index == 1
+
+    def test_model_idx_two_tags_packet(self) -> None:
+        """Boundary: third model in a 3-model run."""
+        emitter, mq = _make_emitter(model_idx=2)
+        emitter.emit(_packet())
+        _key, tagged = mq.get_nowait()
+        assert tagged.placement.model_index == 2
+
+
+# ---------------------------------------------------------------------------
+# Queue key
+# ---------------------------------------------------------------------------
+
+
+class TestEmitterQueueKey:
+    def test_key_equals_model_idx(self) -> None:
+        """Drain loop uses the key to route packets; it must match model_idx."""
+        emitter, mq = _make_emitter(model_idx=2)
+        emitter.emit(_packet())
+        key, _ = mq.get_nowait()
+        assert key == 2
+
+    def test_n1_key_is_zero(self) -> None:
+        emitter, mq = _make_emitter(model_idx=0)
+        emitter.emit(_packet())
+        key, _ = mq.get_nowait()
+        assert key == 0
+
+
+# ---------------------------------------------------------------------------
+# Placement field preservation
+# ---------------------------------------------------------------------------
+
+
+class TestEmitterPlacementPreservation:
+    def test_turn_index_is_preserved(self) -> None:
+        emitter, mq = _make_emitter()
+        emitter.emit(_packet(turn_index=5))
+        _, tagged = mq.get_nowait()
+        assert tagged.placement.turn_index == 5
+
+    def test_tab_index_is_preserved(self) -> None:
+        emitter, mq = _make_emitter()
+        emitter.emit(_packet(tab_index=3))
+        _, tagged = mq.get_nowait()
+        assert tagged.placement.tab_index == 3
+
+    def test_sub_turn_index_is_preserved(self) -> None:
+        emitter, mq = _make_emitter()
+        emitter.emit(_packet(sub_turn_index=2))
+        _, tagged = mq.get_nowait()
+        assert tagged.placement.sub_turn_index == 2
+
+    def test_sub_turn_index_none_is_preserved(self) -> None:
+        emitter, mq = _make_emitter()
+        emitter.emit(_packet(sub_turn_index=None))
+        _, tagged = mq.get_nowait()
+        assert tagged.placement.sub_turn_index is None
+
+    def test_packet_obj_is_not_modified(self) -> None:
+        """The payload object must survive tagging untouched."""
+        emitter, mq = _make_emitter()
+        original_obj = OverallStop(stop_reason="sentinel")
+        pkt = Packet(placement=_placement(), obj=original_obj)
+        emitter.emit(pkt)
+        _, tagged = mq.get_nowait()
+        assert tagged.obj is original_obj
+
+    def test_different_obj_types_are_handled(self) -> None:
+        """Any valid PacketObj type passes through correctly."""
+        emitter, mq = _make_emitter()
+        pkt = Packet(placement=_placement(), obj=ReasoningStart())
+        emitter.emit(pkt)
+        _, tagged = mq.get_nowait()
+        assert isinstance(tagged.obj, ReasoningStart)

backend/tests/unit/onyx/chat/test_multi_model_streaming.py

@@ -1,207 +0,0 @@
-"""Unit tests for multi-model streaming validation and DB helpers.
-
-These are pure unit tests — no real database or LLM calls required.
-The validation logic in run_multi_model_stream fires before any external
-calls, so we can trigger it with lightweight mocks.
-"""
-
-from typing import Any
-from unittest.mock import MagicMock
-from uuid import uuid4
-
-import pytest
-
-from onyx.configs.constants import MessageType
-from onyx.db.chat import set_preferred_response
-from onyx.llm.override_models import LLMOverride
-from onyx.server.query_and_chat.models import SendMessageRequest
-
-
-# ---------------------------------------------------------------------------
-# Helpers
-# ---------------------------------------------------------------------------
-
-
-def _make_request(**kwargs: Any) -> SendMessageRequest:
-    defaults: dict[str, Any] = {
-        "message": "hello",
-        "chat_session_id": uuid4(),
-    }
-    defaults.update(kwargs)
-    return SendMessageRequest(**defaults)
-
-
-def _make_override(provider: str = "openai", version: str = "gpt-4") -> LLMOverride:
-    return LLMOverride(model_provider=provider, model_version=version)
-
-
-def _start_stream(req: SendMessageRequest, overrides: list[LLMOverride]) -> None:
-    """Advance the generator one step to trigger early validation."""
-    from onyx.chat.process_message import run_multi_model_stream
-
-    user = MagicMock()
-    user.is_anonymous = False
-    user.email = "test@example.com"
-    db = MagicMock()
-
-    gen = run_multi_model_stream(req, user, db, overrides)
-    # Calling next() executes until the first yield OR raises.
-    # Validation errors are raised before any yield.
-    next(gen)
-
-
-# ---------------------------------------------------------------------------
-# run_multi_model_stream — validation
-# ---------------------------------------------------------------------------
-
-
-class TestRunMultiModelStreamValidation:
-    def test_single_override_raises(self) -> None:
-        """Exactly 1 override is not multi-model — must raise."""
-        req = _make_request()
-        with pytest.raises(ValueError, match="2-3"):
-            _start_stream(req, [_make_override()])
-
-    def test_four_overrides_raises(self) -> None:
-        """4 overrides exceeds maximum — must raise."""
-        req = _make_request()
-        with pytest.raises(ValueError, match="2-3"):
-            _start_stream(
-                req,
-                [
-                    _make_override("openai", "gpt-4"),
-                    _make_override("anthropic", "claude-3"),
-                    _make_override("google", "gemini-pro"),
-                    _make_override("cohere", "command-r"),
-                ],
-            )
-
-    def test_zero_overrides_raises(self) -> None:
-        """Empty override list raises."""
-        req = _make_request()
-        with pytest.raises(ValueError, match="2-3"):
-            _start_stream(req, [])
-
-    def test_deep_research_raises(self) -> None:
-        """deep_research=True is incompatible with multi-model."""
-        req = _make_request(deep_research=True)
-        with pytest.raises(ValueError, match="not supported"):
-            _start_stream(
-                req, [_make_override(), _make_override("anthropic", "claude-3")]
-            )
-
-    def test_exactly_two_overrides_is_minimum(self) -> None:
-        """Boundary: 1 override fails, 2 passes — ensures fence-post is correct."""
-        req = _make_request()
-        # 1 override must fail
-        with pytest.raises(ValueError, match="2-3"):
-            _start_stream(req, [_make_override()])
-        # 2 overrides must NOT raise ValueError (may raise later due to missing session, that's OK)
-        try:
-            _start_stream(
-                req, [_make_override(), _make_override("anthropic", "claude-3")]
-            )
-        except ValueError as exc:
-            pytest.fail(f"2 overrides should pass validation, got ValueError: {exc}")
-        except Exception:
-            pass  # Any other error means validation passed
-
-
-# ---------------------------------------------------------------------------
-# set_preferred_response — validation (mocked db)
-# ---------------------------------------------------------------------------
-
-
-class TestSetPreferredResponseValidation:
-    def test_user_message_not_found(self) -> None:
-        db = MagicMock()
-        db.get.return_value = None
-
-        with pytest.raises(ValueError, match="not found"):
-            set_preferred_response(
-                db, user_message_id=999, preferred_assistant_message_id=1
-            )
-
-    def test_wrong_message_type(self) -> None:
-        """Cannot set preferred response on a non-USER message."""
-        db = MagicMock()
-        user_msg = MagicMock()
-        user_msg.message_type = MessageType.ASSISTANT  # wrong type
-
-        db.get.return_value = user_msg
-
-        with pytest.raises(ValueError, match="not a user message"):
-            set_preferred_response(
-                db, user_message_id=1, preferred_assistant_message_id=2
-            )
-
-    def test_assistant_message_not_found(self) -> None:
-        db = MagicMock()
-        user_msg = MagicMock()
-        user_msg.message_type = MessageType.USER
-
-        # First call returns user_msg, second call (for assistant) returns None
-        db.get.side_effect = [user_msg, None]
-
-        with pytest.raises(ValueError, match="not found"):
-            set_preferred_response(
-                db, user_message_id=1, preferred_assistant_message_id=2
-            )
-
-    def test_assistant_not_child_of_user(self) -> None:
-        db = MagicMock()
-        user_msg = MagicMock()
-        user_msg.message_type = MessageType.USER
-
-        assistant_msg = MagicMock()
-        assistant_msg.parent_message_id = 999  # different parent
-
-        db.get.side_effect = [user_msg, assistant_msg]
-
-        with pytest.raises(ValueError, match="not a child"):
-            set_preferred_response(
-                db, user_message_id=1, preferred_assistant_message_id=2
-            )
-
-    def test_valid_call_sets_preferred_response_id(self) -> None:
-        db = MagicMock()
-        user_msg = MagicMock()
-        user_msg.message_type = MessageType.USER
-
-        assistant_msg = MagicMock()
-        assistant_msg.parent_message_id = 1  # correct parent
-
-        db.get.side_effect = [user_msg, assistant_msg]
-
-        set_preferred_response(db, user_message_id=1, preferred_assistant_message_id=2)
-
-        assert user_msg.preferred_response_id == 2
-        assert user_msg.latest_child_message_id == 2
-
-
-# ---------------------------------------------------------------------------
-# LLMOverride — display_name field
-# ---------------------------------------------------------------------------
-
-
-class TestLLMOverrideDisplayName:
-    def test_display_name_defaults_none(self) -> None:
-        override = LLMOverride(model_provider="openai", model_version="gpt-4")
-        assert override.display_name is None
-
-    def test_display_name_set(self) -> None:
-        override = LLMOverride(
-            model_provider="openai",
-            model_version="gpt-4",
-            display_name="GPT-4 Turbo",
-        )
-        assert override.display_name == "GPT-4 Turbo"
-
-    def test_display_name_serializes(self) -> None:
-        override = LLMOverride(
-            model_provider="anthropic",
-            model_version="claude-opus-4-6",
-            display_name="Claude Opus",
-        )
-        d = override.model_dump()
-        assert d["display_name"] == "Claude Opus"

backend/tests/unit/onyx/connectors/canvas/test_canvas_connector.py

@@ -0,0 +1,381 @@
+"""Tests for Canvas connector — client (PR1)."""
+
+from typing import Any
+from unittest.mock import MagicMock
+from unittest.mock import patch
+
+import pytest
+
+from onyx.connectors.canvas.client import CanvasApiClient
+from onyx.error_handling.exceptions import OnyxError
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+FAKE_BASE_URL = "https://myschool.instructure.com"
+FAKE_TOKEN = "fake-canvas-token"
+
+
+def _mock_response(
+    status_code: int = 200,
+    json_data: Any = None,
+    link_header: str = "",
+) -> MagicMock:
+    """Create a mock HTTP response with status, json, and Link header."""
+    resp = MagicMock()
+    resp.status_code = status_code
+    resp.reason = "OK" if status_code < 300 else "Error"
+    resp.json.return_value = json_data if json_data is not None else []
+    resp.headers = {"Link": link_header}
+    return resp
+
+
+# ---------------------------------------------------------------------------
+# CanvasApiClient.__init__ tests
+# ---------------------------------------------------------------------------
+
+
+class TestCanvasApiClientInit:
+    def test_success(self) -> None:
+        client = CanvasApiClient(
+            bearer_token=FAKE_TOKEN,
+            canvas_base_url=FAKE_BASE_URL,
+        )
+
+        expected_base_url = f"{FAKE_BASE_URL}/api/v1"
+        expected_host = "myschool.instructure.com"
+
+        assert client.base_url == expected_base_url
+        assert client._expected_host == expected_host
+
+    def test_normalizes_trailing_slash(self) -> None:
+        client = CanvasApiClient(
+            bearer_token=FAKE_TOKEN,
+            canvas_base_url=f"{FAKE_BASE_URL}/",
+        )
+
+        expected_base_url = f"{FAKE_BASE_URL}/api/v1"
+
+        assert client.base_url == expected_base_url
+
+    def test_normalizes_existing_api_v1(self) -> None:
+        client = CanvasApiClient(
+            bearer_token=FAKE_TOKEN,
+            canvas_base_url=f"{FAKE_BASE_URL}/api/v1",
+        )
+
+        expected_base_url = f"{FAKE_BASE_URL}/api/v1"
+
+        assert client.base_url == expected_base_url
+
+    def test_rejects_non_https_scheme(self) -> None:
+        with pytest.raises(ValueError, match="must use https"):
+            CanvasApiClient(
+                bearer_token=FAKE_TOKEN,
+                canvas_base_url="ftp://myschool.instructure.com",
+            )
+
+    def test_rejects_http(self) -> None:
+        with pytest.raises(ValueError, match="must use https"):
+            CanvasApiClient(
+                bearer_token=FAKE_TOKEN,
+                canvas_base_url="http://myschool.instructure.com",
+            )
+
+    def test_rejects_missing_host(self) -> None:
+        with pytest.raises(ValueError, match="must include a valid host"):
+            CanvasApiClient(
+                bearer_token=FAKE_TOKEN,
+                canvas_base_url="https://",
+            )
+
+
+# ---------------------------------------------------------------------------
+# CanvasApiClient._build_url tests
+# ---------------------------------------------------------------------------
+
+
+class TestBuildUrl:
+    def setup_method(self) -> None:
+        self.client = CanvasApiClient(
+            bearer_token=FAKE_TOKEN,
+            canvas_base_url=FAKE_BASE_URL,
+        )
+
+    def test_appends_endpoint(self) -> None:
+        result = self.client._build_url("courses")
+        expected = f"{FAKE_BASE_URL}/api/v1/courses"
+
+        assert result == expected
+
+    def test_strips_leading_slash_from_endpoint(self) -> None:
+        result = self.client._build_url("/courses")
+        expected = f"{FAKE_BASE_URL}/api/v1/courses"
+
+        assert result == expected
+
+
+# ---------------------------------------------------------------------------
+# CanvasApiClient._build_headers tests
+# ---------------------------------------------------------------------------
+
+
+class TestBuildHeaders:
+    def setup_method(self) -> None:
+        self.client = CanvasApiClient(
+            bearer_token=FAKE_TOKEN,
+            canvas_base_url=FAKE_BASE_URL,
+        )
+
+    def test_returns_bearer_auth(self) -> None:
+        result = self.client._build_headers()
+        expected = {"Authorization": f"Bearer {FAKE_TOKEN}"}
+
+        assert result == expected
+
+
+# ---------------------------------------------------------------------------
+# CanvasApiClient.get tests
+# ---------------------------------------------------------------------------
+
+
+class TestGet:
+    def setup_method(self) -> None:
+        self.client = CanvasApiClient(
+            bearer_token=FAKE_TOKEN,
+            canvas_base_url=FAKE_BASE_URL,
+        )
+
+    @patch("onyx.connectors.canvas.client.rl_requests")
+    def test_success_returns_json_and_next_url(self, mock_requests: MagicMock) -> None:
+        next_link = f"<{FAKE_BASE_URL}/api/v1/courses?page=2>; " 'rel="next"'
+        mock_requests.get.return_value = _mock_response(
+            json_data=[{"id": 1}], link_header=next_link
+        )
+
+        data, next_url = self.client.get("courses")
+
+        expected_data = [{"id": 1}]
+        expected_next = f"{FAKE_BASE_URL}/api/v1/courses?page=2"
+
+        assert data == expected_data
+        assert next_url == expected_next
+
+    @patch("onyx.connectors.canvas.client.rl_requests")
+    def test_success_no_next_page(self, mock_requests: MagicMock) -> None:
+        mock_requests.get.return_value = _mock_response(json_data=[{"id": 1}])
+
+        data, next_url = self.client.get("courses")
+
+        assert data == [{"id": 1}]
+        assert next_url is None
+
+    @patch("onyx.connectors.canvas.client.rl_requests")
+    def test_raises_on_error_status(self, mock_requests: MagicMock) -> None:
+        mock_requests.get.return_value = _mock_response(403, {})
+
+        with pytest.raises(OnyxError) as exc_info:
+            self.client.get("courses")
+
+        assert exc_info.value.status_code == 403
+
+    @patch("onyx.connectors.canvas.client.rl_requests")
+    def test_raises_on_404(self, mock_requests: MagicMock) -> None:
+        mock_requests.get.return_value = _mock_response(404, {})
+
+        with pytest.raises(OnyxError) as exc_info:
+            self.client.get("courses")
+
+        assert exc_info.value.status_code == 404
+
+    @patch("onyx.connectors.canvas.client.rl_requests")
+    def test_raises_on_429(self, mock_requests: MagicMock) -> None:
+        mock_requests.get.return_value = _mock_response(429, {})
+
+        with pytest.raises(OnyxError) as exc_info:
+            self.client.get("courses")
+
+        assert exc_info.value.status_code == 429
+
+    @patch("onyx.connectors.canvas.client.rl_requests")
+    def test_skips_params_when_using_full_url(self, mock_requests: MagicMock) -> None:
+        mock_requests.get.return_value = _mock_response(json_data=[])
+        full = f"{FAKE_BASE_URL}/api/v1/courses?page=2"
+
+        self.client.get(params={"per_page": "100"}, full_url=full)
+
+        _, kwargs = mock_requests.get.call_args
+        assert kwargs["params"] is None
+
+    @patch("onyx.connectors.canvas.client.rl_requests")
+    def test_error_extracts_message_from_error_dict(
+        self, mock_requests: MagicMock
+    ) -> None:
+        """Shape 1: {"error": {"message": "Not authorized"}}"""
+        mock_requests.get.return_value = _mock_response(
+            403, {"error": {"message": "Not authorized"}}
+        )
+
+        with pytest.raises(OnyxError) as exc_info:
+            self.client.get("courses")
+
+        result = exc_info.value.detail
+        expected = "Not authorized"
+
+        assert result == expected
+
+    @patch("onyx.connectors.canvas.client.rl_requests")
+    def test_error_extracts_message_from_error_string(
+        self, mock_requests: MagicMock
+    ) -> None:
+        """Shape 2: {"error": "Invalid access token"}"""
+        mock_requests.get.return_value = _mock_response(
+            401, {"error": "Invalid access token"}
+        )
+
+        with pytest.raises(OnyxError) as exc_info:
+            self.client.get("courses")
+
+        result = exc_info.value.detail
+        expected = "Invalid access token"
+
+        assert result == expected
+
+    @patch("onyx.connectors.canvas.client.rl_requests")
+    def test_error_extracts_message_from_errors_list(
+        self, mock_requests: MagicMock
+    ) -> None:
+        """Shape 3: {"errors": [{"message": "Invalid query"}]}"""
+        mock_requests.get.return_value = _mock_response(
+            400, {"errors": [{"message": "Invalid query"}]}
+        )
+
+        with pytest.raises(OnyxError) as exc_info:
+            self.client.get("courses")
+
+        result = exc_info.value.detail
+        expected = "Invalid query"
+
+        assert result == expected
+
+    @patch("onyx.connectors.canvas.client.rl_requests")
+    def test_error_dict_takes_priority_over_errors_list(
+        self, mock_requests: MagicMock
+    ) -> None:
+        """When both error shapes are present, error dict wins."""
+        mock_requests.get.return_value = _mock_response(
+            403, {"error": "Specific error", "errors": [{"message": "Generic"}]}
+        )
+
+        with pytest.raises(OnyxError) as exc_info:
+            self.client.get("courses")
+
+        result = exc_info.value.detail
+        expected = "Specific error"
+
+        assert result == expected
+
+    @patch("onyx.connectors.canvas.client.rl_requests")
+    def test_error_falls_back_to_reason_when_no_json_message(
+        self, mock_requests: MagicMock
+    ) -> None:
+        """Empty error body falls back to response.reason."""
+        mock_requests.get.return_value = _mock_response(500, {})
+
+        with pytest.raises(OnyxError) as exc_info:
+            self.client.get("courses")
+
+        result = exc_info.value.detail
+        expected = "Error"  # from _mock_response's reason for >= 300
+
+        assert result == expected
+
+    @patch("onyx.connectors.canvas.client.rl_requests")
+    def test_invalid_json_on_success_raises(self, mock_requests: MagicMock) -> None:
+        """Invalid JSON on a 2xx response raises OnyxError."""
+        resp = MagicMock()
+        resp.status_code = 200
+        resp.json.side_effect = ValueError("No JSON")
+        resp.headers = {"Link": ""}
+        mock_requests.get.return_value = resp
+
+        with pytest.raises(OnyxError, match="Invalid JSON"):
+            self.client.get("courses")
+
+    @patch("onyx.connectors.canvas.client.rl_requests")
+    def test_invalid_json_on_error_falls_back_to_reason(
+        self, mock_requests: MagicMock
+    ) -> None:
+        """Invalid JSON on a 4xx response falls back to response.reason."""
+        resp = MagicMock()
+        resp.status_code = 500
+        resp.reason = "Internal Server Error"
+        resp.json.side_effect = ValueError("No JSON")
+        resp.headers = {"Link": ""}
+        mock_requests.get.return_value = resp
+
+        with pytest.raises(OnyxError) as exc_info:
+            self.client.get("courses")
+
+        result = exc_info.value.detail
+        expected = "Internal Server Error"
+
+        assert result == expected
+
+
+# ---------------------------------------------------------------------------
+# CanvasApiClient._parse_next_link tests
+# ---------------------------------------------------------------------------
+
+
+class TestParseNextLink:
+    def setup_method(self) -> None:
+        self.client = CanvasApiClient(
+            bearer_token=FAKE_TOKEN,
+            canvas_base_url="https://canvas.example.com",
+        )
+
+    def test_found(self) -> None:
+        header = '<https://canvas.example.com/api/v1/courses?page=2>; rel="next"'
+
+        result = self.client._parse_next_link(header)
+        expected = "https://canvas.example.com/api/v1/courses?page=2"
+
+        assert result == expected
+
+    def test_not_found(self) -> None:
+        header = '<https://canvas.example.com/api/v1/courses?page=1>; rel="current"'
+
+        result = self.client._parse_next_link(header)
+
+        assert result is None
+
+    def test_empty(self) -> None:
+        result = self.client._parse_next_link("")
+
+        assert result is None
+
+    def test_multiple_rels(self) -> None:
+        header = (
+            '<https://canvas.example.com/api/v1/courses?page=1>; rel="current", '
+            '<https://canvas.example.com/api/v1/courses?page=2>; rel="next"'
+        )
+
+        result = self.client._parse_next_link(header)
+        expected = "https://canvas.example.com/api/v1/courses?page=2"
+
+        assert result == expected
+
+    def test_rejects_host_mismatch(self) -> None:
+        header = '<https://evil.example.com/api/v1/courses?page=2>; rel="next"'
+
+        with pytest.raises(OnyxError, match="unexpected host"):
+            self.client._parse_next_link(header)
+
+    def test_rejects_non_https_link(self) -> None:
+        header = '<http://canvas.example.com/api/v1/courses?page=2>; rel="next"'
+
+        with pytest.raises(OnyxError, match="must use https"):
+            self.client._parse_next_link(header)

backend/tests/unit/onyx/connectors/jira/test_jira_bulk_fetch.py

@@ -0,0 +1,147 @@
+from typing import Any
+from unittest.mock import MagicMock
+
+import pytest
+import requests
+from jira import JIRA
+from jira.resources import Issue
+
+from onyx.connectors.jira.connector import bulk_fetch_issues
+
+
+def _make_raw_issue(issue_id: str) -> dict[str, Any]:
+    return {
+        "id": issue_id,
+        "key": f"TEST-{issue_id}",
+        "fields": {"summary": f"Issue {issue_id}"},
+    }
+
+
+def _mock_jira_client() -> MagicMock:
+    mock = MagicMock(spec=JIRA)
+    mock._options = {"server": "https://jira.example.com"}
+    mock._session = MagicMock()
+    mock._get_url = MagicMock(
+        return_value="https://jira.example.com/rest/api/3/issue/bulkfetch"
+    )
+    return mock
+
+
+def test_bulk_fetch_success() -> None:
+    """Happy path: all issues fetched in one request."""
+    client = _mock_jira_client()
+    raw = [_make_raw_issue("1"), _make_raw_issue("2"), _make_raw_issue("3")]
+    resp = MagicMock()
+    resp.json.return_value = {"issues": raw}
+    client._session.post.return_value = resp
+
+    result = bulk_fetch_issues(client, ["1", "2", "3"])
+    assert len(result) == 3
+    assert all(isinstance(r, Issue) for r in result)
+    client._session.post.assert_called_once()
+
+
+def test_bulk_fetch_splits_on_json_error() -> None:
+    """When the full batch fails with JSONDecodeError, sub-batches succeed."""
+    client = _mock_jira_client()
+
+    call_count = 0
+
+    def _post_side_effect(url: str, json: dict[str, Any]) -> MagicMock:  # noqa: ARG001
+        nonlocal call_count
+        call_count += 1
+        ids = json["issueIdsOrKeys"]
+        if len(ids) > 2:
+            resp = MagicMock()
+            resp.json.side_effect = requests.exceptions.JSONDecodeError(
+                "Expecting ',' delimiter", "doc", 2294125
+            )
+            return resp
+
+        resp = MagicMock()
+        resp.json.return_value = {"issues": [_make_raw_issue(i) for i in ids]}
+        return resp
+
+    client._session.post.side_effect = _post_side_effect
+
+    result = bulk_fetch_issues(client, ["1", "2", "3", "4"])
+    assert len(result) == 4
+    returned_ids = {r.raw["id"] for r in result}
+    assert returned_ids == {"1", "2", "3", "4"}
+    assert call_count > 1
+
+
+def test_bulk_fetch_raises_on_single_unfetchable_issue() -> None:
+    """A single issue that always fails JSON decode raises after splitting."""
+    client = _mock_jira_client()
+
+    def _post_side_effect(url: str, json: dict[str, Any]) -> MagicMock:  # noqa: ARG001
+        ids = json["issueIdsOrKeys"]
+        if "bad" in ids:
+            resp = MagicMock()
+            resp.json.side_effect = requests.exceptions.JSONDecodeError(
+                "Expecting ',' delimiter", "doc", 100
+            )
+            return resp
+
+        resp = MagicMock()
+        resp.json.return_value = {"issues": [_make_raw_issue(i) for i in ids]}
+        return resp
+
+    client._session.post.side_effect = _post_side_effect
+
+    with pytest.raises(requests.exceptions.JSONDecodeError):
+        bulk_fetch_issues(client, ["1", "bad", "2"])
+
+
+def test_bulk_fetch_non_json_error_propagates() -> None:
+    """Non-JSONDecodeError exceptions still propagate."""
+    client = _mock_jira_client()
+
+    resp = MagicMock()
+    resp.json.side_effect = ValueError("something else broke")
+    client._session.post.return_value = resp
+
+    try:
+        bulk_fetch_issues(client, ["1"])
+        assert False, "Expected ValueError to propagate"
+    except ValueError:
+        pass
+
+
+def test_bulk_fetch_with_fields() -> None:
+    """Fields parameter is forwarded correctly."""
+    client = _mock_jira_client()
+    raw = [_make_raw_issue("1")]
+    resp = MagicMock()
+    resp.json.return_value = {"issues": raw}
+    client._session.post.return_value = resp
+
+    bulk_fetch_issues(client, ["1"], fields="summary,description")
+
+    call_payload = client._session.post.call_args[1]["json"]
+    assert call_payload["fields"] == ["summary", "description"]
+
+
+def test_bulk_fetch_recursive_splitting_raises_on_bad_issue() -> None:
+    """With a 6-issue batch where one is bad, recursion isolates it and raises."""
+    client = _mock_jira_client()
+    bad_id = "BAD"
+
+    def _post_side_effect(url: str, json: dict[str, Any]) -> MagicMock:  # noqa: ARG001
+        ids = json["issueIdsOrKeys"]
+        if bad_id in ids:
+            resp = MagicMock()
+            resp.json.side_effect = requests.exceptions.JSONDecodeError(
+                "truncated", "doc", 999
+            )
+            return resp
+
+        resp = MagicMock()
+        resp.json.return_value = {"issues": [_make_raw_issue(i) for i in ids]}
+        return resp
+
+    client._session.post.side_effect = _post_side_effect
+
+    with pytest.raises(requests.exceptions.JSONDecodeError):
+        bulk_fetch_issues(client, ["1", "2", bad_id, "3", "4", "5"])

backend/tests/unit/onyx/connectors/jira/test_jira_permission_sync.py

@@ -1,3 +1,5 @@
+from datetime import datetime
+from datetime import timezone
 from unittest.mock import MagicMock
 from unittest.mock import patch
 
@@ -31,6 +33,7 @@ def mock_jira_cc_pair(
         "jira_base_url": jira_base_url,
         "project_key": project_key,
     }
+    mock_cc_pair.connector.indexing_start = None
 
     return mock_cc_pair
 
@@ -65,3 +68,75 @@ def test_jira_permission_sync(
             fetch_all_existing_docs_ids_fn=mock_fetch_all_existing_docs_ids_fn,
         ):
             print(doc)
+
+
+def test_jira_doc_sync_passes_indexing_start(
+    jira_connector: JiraConnector,
+    mock_jira_cc_pair: MagicMock,
+    mock_fetch_all_existing_docs_fn: MagicMock,
+    mock_fetch_all_existing_docs_ids_fn: MagicMock,
+) -> None:
+    """Verify that generic_doc_sync derives indexing_start from cc_pair
+    and forwards it to retrieve_all_slim_docs_perm_sync."""
+    indexing_start_dt = datetime(2025, 6, 1, tzinfo=timezone.utc)
+    mock_jira_cc_pair.connector.indexing_start = indexing_start_dt
+
+    with patch("onyx.connectors.jira.connector.build_jira_client") as mock_build_client:
+        mock_build_client.return_value = jira_connector._jira_client
+        assert jira_connector._jira_client is not None
+        jira_connector._jira_client._options = MagicMock()
+        jira_connector._jira_client._options.return_value = {
+            "rest_api_version": JIRA_SERVER_API_VERSION
+        }
+
+        with patch.object(
+            type(jira_connector),
+            "retrieve_all_slim_docs_perm_sync",
+            return_value=iter([]),
+        ) as mock_retrieve:
+            list(
+                jira_doc_sync(
+                    cc_pair=mock_jira_cc_pair,
+                    fetch_all_existing_docs_fn=mock_fetch_all_existing_docs_fn,
+                    fetch_all_existing_docs_ids_fn=mock_fetch_all_existing_docs_ids_fn,
+                )
+            )
+
+            mock_retrieve.assert_called_once()
+            call_kwargs = mock_retrieve.call_args
+            assert call_kwargs.kwargs["start"] == indexing_start_dt.timestamp()
+
+
+def test_jira_doc_sync_passes_none_when_no_indexing_start(
+    jira_connector: JiraConnector,
+    mock_jira_cc_pair: MagicMock,
+    mock_fetch_all_existing_docs_fn: MagicMock,
+    mock_fetch_all_existing_docs_ids_fn: MagicMock,
+) -> None:
+    """Verify that indexing_start is None when the connector has no indexing_start set."""
+    mock_jira_cc_pair.connector.indexing_start = None
+
+    with patch("onyx.connectors.jira.connector.build_jira_client") as mock_build_client:
+        mock_build_client.return_value = jira_connector._jira_client
+        assert jira_connector._jira_client is not None
+        jira_connector._jira_client._options = MagicMock()
+        jira_connector._jira_client._options.return_value = {
+            "rest_api_version": JIRA_SERVER_API_VERSION
+        }
+
+        with patch.object(
+            type(jira_connector),
+            "retrieve_all_slim_docs_perm_sync",
+            return_value=iter([]),
+        ) as mock_retrieve:
+            list(
+                jira_doc_sync(
+                    cc_pair=mock_jira_cc_pair,
+                    fetch_all_existing_docs_fn=mock_fetch_all_existing_docs_fn,
+                    fetch_all_existing_docs_ids_fn=mock_fetch_all_existing_docs_ids_fn,
+                )
+            )
+
+            mock_retrieve.assert_called_once()
+            call_kwargs = mock_retrieve.call_args
+            assert call_kwargs.kwargs["start"] is None

backend/tests/unit/onyx/db/test_voice.py

@@ -272,13 +272,13 @@ class TestUpsertVoiceProvider:
 class TestDeleteVoiceProvider:
     """Tests for delete_voice_provider."""
 
-    def test_soft_deletes_provider_when_found(self, mock_db_session: MagicMock) -> None:
+    def test_hard_deletes_provider_when_found(self, mock_db_session: MagicMock) -> None:
         provider = _make_voice_provider(id=1)
         mock_db_session.scalar.return_value = provider
 
         delete_voice_provider(mock_db_session, 1)
 
-        assert provider.deleted is True
+        mock_db_session.delete.assert_called_once_with(provider)
         mock_db_session.flush.assert_called_once()
 
     def test_does_nothing_when_provider_not_found(

backend/tests/unit/onyx/document_index/opensearch/test_opensearch_batch_flush.py

@@ -0,0 +1,223 @@
+from unittest.mock import MagicMock
+from unittest.mock import patch
+
+from onyx.access.models import DocumentAccess
+from onyx.configs.constants import DocumentSource
+from onyx.connectors.models import Document
+from onyx.connectors.models import TextSection
+from onyx.document_index.interfaces_new import IndexingMetadata
+from onyx.document_index.interfaces_new import TenantState
+from onyx.document_index.opensearch.opensearch_document_index import (
+    OpenSearchDocumentIndex,
+)
+from onyx.indexing.models import ChunkEmbedding
+from onyx.indexing.models import DocMetadataAwareIndexChunk
+
+
+def _make_chunk(
+    doc_id: str,
+    chunk_id: int,
+) -> DocMetadataAwareIndexChunk:
+    """Creates a minimal DocMetadataAwareIndexChunk for testing."""
+    doc = Document(
+        id=doc_id,
+        sections=[TextSection(text="test", link="http://test.com")],
+        source=DocumentSource.FILE,
+        semantic_identifier="test_doc",
+        metadata={},
+    )
+    access = DocumentAccess.build(
+        user_emails=[],
+        user_groups=[],
+        external_user_emails=[],
+        external_user_group_ids=[],
+        is_public=True,
+    )
+    return DocMetadataAwareIndexChunk(
+        chunk_id=chunk_id,
+        blurb="test",
+        content="test content",
+        source_links={0: "http://test.com"},
+        image_file_id=None,
+        section_continuation=False,
+        source_document=doc,
+        title_prefix="",
+        metadata_suffix_semantic="",
+        metadata_suffix_keyword="",
+        mini_chunk_texts=None,
+        large_chunk_id=None,
+        doc_summary="",
+        chunk_context="",
+        contextual_rag_reserved_tokens=0,
+        embeddings=ChunkEmbedding(full_embedding=[0.1] * 10, mini_chunk_embeddings=[]),
+        title_embedding=[0.1] * 10,
+        tenant_id="test_tenant",
+        access=access,
+        document_sets=set(),
+        user_project=[],
+        personas=[],
+        boost=0,
+        aggregated_chunk_boost_factor=1.0,
+        ancestor_hierarchy_node_ids=[],
+    )
+
+
+def _make_index() -> tuple[OpenSearchDocumentIndex, MagicMock]:
+    """Creates an OpenSearchDocumentIndex with a mocked client.
+    Returns the index and the mock for bulk_index_documents."""
+    mock_client = MagicMock()
+    mock_bulk = MagicMock()
+    mock_client.bulk_index_documents = mock_bulk
+
+    tenant_state = TenantState(tenant_id="test_tenant", multitenant=False)
+
+    index = OpenSearchDocumentIndex.__new__(OpenSearchDocumentIndex)
+    index._index_name = "test_index"
+    index._client = mock_client
+    index._tenant_state = tenant_state
+
+    return index, mock_bulk
+
+
+def _make_metadata(doc_id: str, chunk_count: int) -> IndexingMetadata:
+    return IndexingMetadata(
+        doc_id_to_chunk_cnt_diff={
+            doc_id: IndexingMetadata.ChunkCounts(
+                old_chunk_cnt=0,
+                new_chunk_cnt=chunk_count,
+            ),
+        },
+    )
+
+
+@patch(
+    "onyx.document_index.opensearch.opensearch_document_index.MAX_CHUNKS_PER_DOC_BATCH",
+    100,
+)
+def test_single_doc_under_batch_limit_flushes_once() -> None:
+    """A document with fewer chunks than MAX_CHUNKS_PER_DOC_BATCH should flush once."""
+    index, mock_bulk = _make_index()
+    doc_id = "doc_1"
+    num_chunks = 50
+    chunks = [_make_chunk(doc_id, i) for i in range(num_chunks)]
+    metadata = _make_metadata(doc_id, num_chunks)
+
+    with patch.object(index, "delete", return_value=0):
+        index.index(chunks, metadata)
+
+    assert mock_bulk.call_count == 1
+    batch_arg = mock_bulk.call_args_list[0]
+    assert len(batch_arg.kwargs["documents"]) == num_chunks
+
+
+@patch(
+    "onyx.document_index.opensearch.opensearch_document_index.MAX_CHUNKS_PER_DOC_BATCH",
+    100,
+)
+def test_single_doc_over_batch_limit_flushes_multiple_times() -> None:
+    """A document with more chunks than MAX_CHUNKS_PER_DOC_BATCH should flush multiple times."""
+    index, mock_bulk = _make_index()
+    doc_id = "doc_1"
+    num_chunks = 250
+    chunks = [_make_chunk(doc_id, i) for i in range(num_chunks)]
+    metadata = _make_metadata(doc_id, num_chunks)
+
+    with patch.object(index, "delete", return_value=0):
+        index.index(chunks, metadata)
+
+    # 250 chunks / 100 per batch = 3 flushes (100 + 100 + 50)
+    assert mock_bulk.call_count == 3
+    batch_sizes = [len(call.kwargs["documents"]) for call in mock_bulk.call_args_list]
+    assert batch_sizes == [100, 100, 50]
+
+
+@patch(
+    "onyx.document_index.opensearch.opensearch_document_index.MAX_CHUNKS_PER_DOC_BATCH",
+    100,
+)
+def test_single_doc_exactly_at_batch_limit() -> None:
+    """A document with exactly MAX_CHUNKS_PER_DOC_BATCH chunks should flush once
+    (the flush happens on the next chunk, not at the boundary)."""
+    index, mock_bulk = _make_index()
+    doc_id = "doc_1"
+    num_chunks = 100
+    chunks = [_make_chunk(doc_id, i) for i in range(num_chunks)]
+    metadata = _make_metadata(doc_id, num_chunks)
+
+    with patch.object(index, "delete", return_value=0):
+        index.index(chunks, metadata)
+
+    # 100 chunks hit the >= check on chunk 101 which doesn't exist,
+    # so final flush handles all 100
+    # Actually: the elif fires when len(current_chunks) >= 100, which happens
+    # when current_chunks has 100 items and the 101st chunk arrives.
+    # With exactly 100 chunks, the 100th chunk makes len == 99, then appended -> 100.
+    # No 101st chunk arrives, so the final flush handles all 100.
+    assert mock_bulk.call_count == 1
+
+
+@patch(
+    "onyx.document_index.opensearch.opensearch_document_index.MAX_CHUNKS_PER_DOC_BATCH",
+    100,
+)
+def test_single_doc_one_over_batch_limit() -> None:
+    """101 chunks for one doc: first 100 flushed when the 101st arrives, then
+    the 101st is flushed at the end."""
+    index, mock_bulk = _make_index()
+    doc_id = "doc_1"
+    num_chunks = 101
+    chunks = [_make_chunk(doc_id, i) for i in range(num_chunks)]
+    metadata = _make_metadata(doc_id, num_chunks)
+
+    with patch.object(index, "delete", return_value=0):
+        index.index(chunks, metadata)
+
+    assert mock_bulk.call_count == 2
+    batch_sizes = [len(call.kwargs["documents"]) for call in mock_bulk.call_args_list]
+    assert batch_sizes == [100, 1]
+
+
+@patch(
+    "onyx.document_index.opensearch.opensearch_document_index.MAX_CHUNKS_PER_DOC_BATCH",
+    100,
+)
+def test_multiple_docs_each_under_limit_flush_per_doc() -> None:
+    """Multiple documents each under the batch limit should flush once per document."""
+    index, mock_bulk = _make_index()
+    chunks = []
+    for doc_idx in range(3):
+        doc_id = f"doc_{doc_idx}"
+        for chunk_idx in range(50):
+            chunks.append(_make_chunk(doc_id, chunk_idx))
+
+    metadata = IndexingMetadata(
+        doc_id_to_chunk_cnt_diff={
+            f"doc_{i}": IndexingMetadata.ChunkCounts(old_chunk_cnt=0, new_chunk_cnt=50)
+            for i in range(3)
+        },
+    )
+
+    with patch.object(index, "delete", return_value=0):
+        index.index(chunks, metadata)
+
+    # 3 documents = 3 flushes (one per doc boundary + final)
+    assert mock_bulk.call_count == 3
+
+
+@patch(
+    "onyx.document_index.opensearch.opensearch_document_index.MAX_CHUNKS_PER_DOC_BATCH",
+    100,
+)
+def test_delete_called_once_per_document() -> None:
+    """Even with multiple flushes for a single document, delete should only be
+    called once per document."""
+    index, _mock_bulk = _make_index()
+    doc_id = "doc_1"
+    num_chunks = 250
+    chunks = [_make_chunk(doc_id, i) for i in range(num_chunks)]
+    metadata = _make_metadata(doc_id, num_chunks)
+
+    with patch.object(index, "delete", return_value=0) as mock_delete:
+        index.index(chunks, metadata)
+
+    mock_delete.assert_called_once_with(doc_id, None)