update error

Co-authored-by: Richard Kuo (Danswer) <rkuo@onyx.app>

.github/workflows/docker-build-push-cloud-web-container-on-tag.yml

@@ -65,6 +65,7 @@ jobs:
             NEXT_PUBLIC_POSTHOG_KEY=${{ secrets.POSTHOG_KEY }}
             NEXT_PUBLIC_POSTHOG_HOST=${{ secrets.POSTHOG_HOST }}
             NEXT_PUBLIC_SENTRY_DSN=${{ secrets.SENTRY_DSN }}
+            NEXT_PUBLIC_STRIPE_PUBLISHABLE_KEY=${{ secrets.STRIPE_PUBLISHABLE_KEY }}
             NEXT_PUBLIC_GTM_ENABLED=true
             NEXT_PUBLIC_FORGOT_PASSWORD_ENABLED=true
             NEXT_PUBLIC_INCLUDE_ERROR_POPUP_SUPPORT_LINK=true

.github/workflows/docker-build-push-model-server-container-on-tag.yml

@@ -12,7 +12,32 @@ env:
   BUILDKIT_PROGRESS: plain
 
 jobs:
+  # 1) Preliminary job to check if the changed files are relevant
+  check_model_server_changes:
+    runs-on: ubuntu-latest
+    outputs:
+      changed: ${{ steps.check.outputs.changed }}
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Check if relevant files changed
+        id: check
+        run: |
+          # Default to "false"
+          echo "changed=false" >> $GITHUB_OUTPUT
+
+          # Compare the previous commit (github.event.before) to the current one (github.sha)
+          # If any file in backend/model_server/** or backend/Dockerfile.model_server is changed,
+          # set changed=true
+          if git diff --name-only ${{ github.event.before }} ${{ github.sha }} \
+             | grep -E '^backend/model_server/|^backend/Dockerfile.model_server'; then
+            echo "changed=true" >> $GITHUB_OUTPUT
+          fi
+
   build-amd64:
+    needs: [check_model_server_changes]
+    if: needs.check_model_server_changes.outputs.changed == 'true'
     runs-on:
       [runs-on, runner=8cpu-linux-x64, "run-id=${{ github.run_id }}-amd64"]
     steps:
@@ -52,6 +77,8 @@ jobs:
           provenance: false
 
   build-arm64:
+    needs: [check_model_server_changes]
+    if: needs.check_model_server_changes.outputs.changed == 'true'
     runs-on:
       [runs-on, runner=8cpu-linux-x64, "run-id=${{ github.run_id }}-arm64"]
     steps:
@@ -91,7 +118,8 @@ jobs:
           provenance: false
 
   merge-and-scan:
-    needs: [build-amd64, build-arm64]
+    needs: [build-amd64, build-arm64, check_model_server_changes]
+    if: needs.check_model_server_changes.outputs.changed == 'true'
     runs-on: ubuntu-latest
     steps:
       - name: Login to Docker Hub

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

@@ -44,6 +44,9 @@ env:
   SHAREPOINT_CLIENT_SECRET: ${{ secrets.SHAREPOINT_CLIENT_SECRET }}
   SHAREPOINT_CLIENT_DIRECTORY_ID: ${{ secrets.SHAREPOINT_CLIENT_DIRECTORY_ID }}
   SHAREPOINT_SITE: ${{ secrets.SHAREPOINT_SITE }}
+  # Gitbook
+  GITBOOK_SPACE_ID: ${{ secrets.GITBOOK_SPACE_ID }}
+  GITBOOK_API_KEY: ${{ secrets.GITBOOK_API_KEY }}
 
 jobs:
   connectors-check:

backend/Dockerfile

@@ -35,7 +35,9 @@ RUN apt-get update && \
         libuuid1=2.38.1-5+deb12u1 \
         libxmlsec1-dev \
         pkg-config \
-        gcc && \
+        gcc \
+        nano \
+        vim && \
     rm -rf /var/lib/apt/lists/* && \
     apt-get clean
 

backend/alembic/versions/2cdeff6d8c93_set_built_in_to_default.py

@@ -0,0 +1,32 @@
+"""set built in to default
+
+Revision ID: 2cdeff6d8c93
+Revises: f5437cc136c5
+Create Date: 2025-02-11 14:57:51.308775
+
+"""
+from alembic import op
+
+
+# revision identifiers, used by Alembic.
+revision = "2cdeff6d8c93"
+down_revision = "f5437cc136c5"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    # Prior to this migration / point in the codebase history,
+    # built in personas were implicitly treated as default personas (with no option to change this)
+    # This migration makes that explicit
+    op.execute(
+        """
+        UPDATE persona
+        SET is_default_persona = TRUE
+        WHERE builtin_persona = TRUE
+    """
+    )
+
+
+def downgrade() -> None:
+    pass

backend/alembic/versions/f39c5794c10a_add_background_errors_table.py

@@ -0,0 +1,40 @@
+"""Add background errors table
+
+Revision ID: f39c5794c10a
+Revises: 2cdeff6d8c93
+Create Date: 2025-02-12 17:11:14.527876
+
+"""
+from alembic import op
+import sqlalchemy as sa
+
+# revision identifiers, used by Alembic.
+revision = "f39c5794c10a"
+down_revision = "2cdeff6d8c93"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        "background_error",
+        sa.Column("id", sa.Integer(), nullable=False),
+        sa.Column("message", sa.String(), nullable=False),
+        sa.Column(
+            "time_created",
+            sa.DateTime(timezone=True),
+            server_default=sa.text("now()"),
+            nullable=False,
+        ),
+        sa.Column("cc_pair_id", sa.Integer(), nullable=True),
+        sa.PrimaryKeyConstraint("id"),
+        sa.ForeignKeyConstraint(
+            ["cc_pair_id"],
+            ["connector_credential_pair.id"],
+            ondelete="CASCADE",
+        ),
+    )
+
+
+def downgrade() -> None:
+    op.drop_table("background_error")

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

@@ -1,44 +1,46 @@
 from datetime import timedelta
 from typing import Any
 
+from onyx.background.celery.tasks.beat_schedule import (
+    beat_cloud_tasks as base_beat_system_tasks,
+)
 from onyx.background.celery.tasks.beat_schedule import BEAT_EXPIRES_DEFAULT
 from onyx.background.celery.tasks.beat_schedule import (
-    cloud_tasks_to_schedule as base_cloud_tasks_to_schedule,
+    beat_task_templates as base_beat_task_templates,
 )
+from onyx.background.celery.tasks.beat_schedule import generate_cloud_tasks
 from onyx.background.celery.tasks.beat_schedule import (
-    tasks_to_schedule as base_tasks_to_schedule,
+    get_tasks_to_schedule as base_get_tasks_to_schedule,
 )
-from onyx.configs.constants import ONYX_CLOUD_CELERY_TASK_PREFIX
 from onyx.configs.constants import OnyxCeleryPriority
 from onyx.configs.constants import OnyxCeleryTask
 from shared_configs.configs import MULTI_TENANT
 
-ee_cloud_tasks_to_schedule = [
-    {
-        "name": f"{ONYX_CLOUD_CELERY_TASK_PREFIX}_autogenerate-usage-report",
-        "task": OnyxCeleryTask.CLOUD_BEAT_TASK_GENERATOR,
-        "schedule": timedelta(days=30),
-        "options": {
-            "priority": OnyxCeleryPriority.HIGHEST,
-            "expires": BEAT_EXPIRES_DEFAULT,
+ee_beat_system_tasks: list[dict] = []
+
+ee_beat_task_templates: list[dict] = []
+ee_beat_task_templates.extend(
+    [
+        {
+            "name": "autogenerate-usage-report",
+            "task": OnyxCeleryTask.AUTOGENERATE_USAGE_REPORT_TASK,
+            "schedule": timedelta(days=30),
+            "options": {
+                "priority": OnyxCeleryPriority.MEDIUM,
+                "expires": BEAT_EXPIRES_DEFAULT,
+            },
         },
-        "kwargs": {
-            "task_name": OnyxCeleryTask.AUTOGENERATE_USAGE_REPORT_TASK,
+        {
+            "name": "check-ttl-management",
+            "task": OnyxCeleryTask.CHECK_TTL_MANAGEMENT_TASK,
+            "schedule": timedelta(hours=1),
+            "options": {
+                "priority": OnyxCeleryPriority.MEDIUM,
+                "expires": BEAT_EXPIRES_DEFAULT,
+            },
         },
-    },
-    {
-        "name": f"{ONYX_CLOUD_CELERY_TASK_PREFIX}_check-ttl-management",
-        "task": OnyxCeleryTask.CLOUD_BEAT_TASK_GENERATOR,
-        "schedule": timedelta(hours=1),
-        "options": {
-            "priority": OnyxCeleryPriority.HIGHEST,
-            "expires": BEAT_EXPIRES_DEFAULT,
-        },
-        "kwargs": {
-            "task_name": OnyxCeleryTask.CHECK_TTL_MANAGEMENT_TASK,
-        },
-    },
-]
+    ]
+)
 
 ee_tasks_to_schedule: list[dict] = []
 
@@ -65,9 +67,14 @@ if not MULTI_TENANT:
     ]
 
 
-def get_cloud_tasks_to_schedule() -> list[dict[str, Any]]:
-    return ee_cloud_tasks_to_schedule + base_cloud_tasks_to_schedule
+def get_cloud_tasks_to_schedule(beat_multiplier: float) -> list[dict[str, Any]]:
+    beat_system_tasks = ee_beat_system_tasks + base_beat_system_tasks
+    beat_task_templates = ee_beat_task_templates + base_beat_task_templates
+    cloud_tasks = generate_cloud_tasks(
+        beat_system_tasks, beat_task_templates, beat_multiplier
+    )
+    return cloud_tasks
 
 
 def get_tasks_to_schedule() -> list[dict[str, Any]]:
-    return ee_tasks_to_schedule + base_tasks_to_schedule
+    return ee_tasks_to_schedule + base_get_tasks_to_schedule()

backend/ee/onyx/configs/app_configs.py

@@ -77,3 +77,5 @@ POSTHOG_HOST = os.environ.get("POSTHOG_HOST") or "https://us.i.posthog.com"
 HUBSPOT_TRACKING_URL = os.environ.get("HUBSPOT_TRACKING_URL")
 
 ANONYMOUS_USER_COOKIE_NAME = "onyx_anonymous_user"
+
+GATED_TENANTS_KEY = "gated_tenants"

backend/ee/onyx/db/persona.py

@@ -15,6 +15,9 @@ def make_persona_private(
     group_ids: list[int] | None,
     db_session: Session,
 ) -> None:
+    """NOTE(rkuo): This function batches all updates into a single commit. If we don't
+    dedupe the inputs, the commit will exception."""
+
     db_session.query(Persona__User).filter(
         Persona__User.persona_id == persona_id
     ).delete(synchronize_session="fetch")
@@ -23,19 +26,22 @@ def make_persona_private(
     ).delete(synchronize_session="fetch")
 
     if user_ids:
-        for user_uuid in user_ids:
-            db_session.add(Persona__User(persona_id=persona_id, user_id=user_uuid))
+        user_ids_set = set(user_ids)
+        for user_id in user_ids_set:
+            db_session.add(Persona__User(persona_id=persona_id, user_id=user_id))
 
             create_notification(
-                user_id=user_uuid,
+                user_id=user_id,
                 notif_type=NotificationType.PERSONA_SHARED,
                 db_session=db_session,
                 additional_data=PersonaSharedNotificationData(
                     persona_id=persona_id,
                 ).model_dump(),
             )
+
     if group_ids:
-        for group_id in group_ids:
+        group_ids_set = set(group_ids)
+        for group_id in group_ids_set:
             db_session.add(
                 Persona__UserGroup(persona_id=persona_id, user_group_id=group_id)
             )

backend/ee/onyx/external_permissions/confluence/group_sync.py

@@ -1,5 +1,6 @@
 from ee.onyx.db.external_perm import ExternalUserGroup
 from ee.onyx.external_permissions.confluence.constants import ALL_CONF_EMAILS_GROUP_NAME
+from onyx.background.error_logging import emit_background_error
 from onyx.connectors.confluence.onyx_confluence import build_confluence_client
 from onyx.connectors.confluence.onyx_confluence import OnyxConfluence
 from onyx.connectors.confluence.utils import get_user_email_from_username__server
@@ -10,7 +11,7 @@ logger = setup_logger()
 
 
 def _build_group_member_email_map(
-    confluence_client: OnyxConfluence,
+    confluence_client: OnyxConfluence, cc_pair_id: int
 ) -> dict[str, set[str]]:
     group_member_emails: dict[str, set[str]] = {}
     for user_result in confluence_client.paginated_cql_user_retrieval():
@@ -18,8 +19,11 @@ def _build_group_member_email_map(
 
         user = user_result.get("user", {})
         if not user:
-            logger.warning(f"user result missing user field: {user_result}")
+            msg = f"user result missing user field: {user_result}"
+            emit_background_error(msg, cc_pair_id=cc_pair_id)
+            logger.error(msg)
             continue
+
         email = user.get("email")
         if not email:
             # This field is only present in Confluence Server
@@ -32,7 +36,12 @@ def _build_group_member_email_map(
                 )
         if not email:
             # If we still don't have an email, skip this user
-            logger.warning(f"user result missing email field: {user_result}")
+            msg = f"user result missing email field: {user_result}"
+            if user.get("type") == "app":
+                logger.warning(msg)
+            else:
+                emit_background_error(msg, cc_pair_id=cc_pair_id)
+                logger.error(msg)
             continue
 
         all_users_groups: set[str] = set()
@@ -42,11 +51,18 @@ def _build_group_member_email_map(
             group_member_emails.setdefault(group_id, set()).add(email)
             all_users_groups.add(group_id)
 
-        if not group_member_emails:
-            logger.warning(f"No groups found for user with email: {email}")
+        if not all_users_groups:
+            msg = f"No groups found for user with email: {email}"
+            emit_background_error(msg, cc_pair_id=cc_pair_id)
+            logger.error(msg)
         else:
             logger.debug(f"Found groups {all_users_groups} for user with email {email}")
 
+    if not group_member_emails:
+        msg = "No groups found for any users."
+        emit_background_error(msg, cc_pair_id=cc_pair_id)
+        logger.error(msg)
+
     return group_member_emails
 
 
@@ -61,6 +77,7 @@ def confluence_group_sync(
 
     group_member_email_map = _build_group_member_email_map(
         confluence_client=confluence_client,
+        cc_pair_id=cc_pair.id,
     )
     onyx_groups: list[ExternalUserGroup] = []
     all_found_emails = set()

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

@@ -83,6 +83,7 @@ def handle_search_request(
         user=user,
         llm=llm,
         fast_llm=fast_llm,
+        skip_query_analysis=False,
         db_session=db_session,
         bypass_acl=False,
     )

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

@@ -18,11 +18,16 @@ from ee.onyx.server.tenants.anonymous_user_path import (
 from ee.onyx.server.tenants.anonymous_user_path import modify_anonymous_user_path
 from ee.onyx.server.tenants.anonymous_user_path import validate_anonymous_user_path
 from ee.onyx.server.tenants.billing import fetch_billing_information
+from ee.onyx.server.tenants.billing import fetch_stripe_checkout_session
 from ee.onyx.server.tenants.billing import fetch_tenant_stripe_information
 from ee.onyx.server.tenants.models import AnonymousUserPath
 from ee.onyx.server.tenants.models import BillingInformation
 from ee.onyx.server.tenants.models import ImpersonateRequest
 from ee.onyx.server.tenants.models import ProductGatingRequest
+from ee.onyx.server.tenants.models import ProductGatingResponse
+from ee.onyx.server.tenants.models import SubscriptionSessionResponse
+from ee.onyx.server.tenants.models import SubscriptionStatusResponse
+from ee.onyx.server.tenants.product_gating import store_product_gating
 from ee.onyx.server.tenants.provisioning import delete_user_from_control_plane
 from ee.onyx.server.tenants.user_mapping import get_tenant_id_for_email
 from ee.onyx.server.tenants.user_mapping import remove_all_users_from_tenant
@@ -39,12 +44,9 @@ from onyx.db.auth import get_user_count
 from onyx.db.engine import get_current_tenant_id
 from onyx.db.engine import get_session
 from onyx.db.engine import get_session_with_tenant
-from onyx.db.notification import create_notification
 from onyx.db.users import delete_user_from_db
 from onyx.db.users import get_user_by_email
 from onyx.server.manage.models import UserByEmail
-from onyx.server.settings.store import load_settings
-from onyx.server.settings.store import store_settings
 from onyx.utils.logger import setup_logger
 from shared_configs.contextvars import CURRENT_TENANT_ID_CONTEXTVAR
 
@@ -126,37 +128,29 @@ async def login_as_anonymous_user(
 @router.post("/product-gating")
 def gate_product(
     product_gating_request: ProductGatingRequest, _: None = Depends(control_plane_dep)
-) -> None:
+) -> ProductGatingResponse:
     """
     Gating the product means that the product is not available to the tenant.
     They will be directed to the billing page.
-    We gate the product when
-    1) User has ended free trial without adding payment method
-    2) User's card has declined
+    We gate the product when their subscription has ended.
     """
-    tenant_id = product_gating_request.tenant_id
-    token = CURRENT_TENANT_ID_CONTEXTVAR.set(tenant_id)
+    try:
+        store_product_gating(
+            product_gating_request.tenant_id, product_gating_request.application_status
+        )
+        return ProductGatingResponse(updated=True, error=None)
 
-    settings = load_settings()
-    settings.product_gating = product_gating_request.product_gating
-    store_settings(settings)
-
-    if product_gating_request.notification:
-        with get_session_with_tenant(tenant_id) as db_session:
-            create_notification(None, product_gating_request.notification, db_session)
-
-    if token is not None:
-        CURRENT_TENANT_ID_CONTEXTVAR.reset(token)
+    except Exception as e:
+        logger.exception("Failed to gate product")
+        return ProductGatingResponse(updated=False, error=str(e))
 
 
-@router.get("/billing-information", response_model=BillingInformation)
+@router.get("/billing-information")
 async def billing_information(
     _: User = Depends(current_admin_user),
-) -> BillingInformation:
+) -> BillingInformation | SubscriptionStatusResponse:
     logger.info("Fetching billing information")
-    return BillingInformation(
-        **fetch_billing_information(CURRENT_TENANT_ID_CONTEXTVAR.get())
-    )
+    return fetch_billing_information(CURRENT_TENANT_ID_CONTEXTVAR.get())
 
 
 @router.post("/create-customer-portal-session")
@@ -169,9 +163,10 @@ async def create_customer_portal_session(_: User = Depends(current_admin_user))
         if not stripe_customer_id:
             raise HTTPException(status_code=400, detail="Stripe customer ID not found")
         logger.info(stripe_customer_id)
+
         portal_session = stripe.billing_portal.Session.create(
             customer=stripe_customer_id,
-            return_url=f"{WEB_DOMAIN}/admin/cloud-settings",
+            return_url=f"{WEB_DOMAIN}/admin/billing",
         )
         logger.info(portal_session)
         return {"url": portal_session.url}
@@ -180,6 +175,20 @@ async def create_customer_portal_session(_: User = Depends(current_admin_user))
         raise HTTPException(status_code=500, detail=str(e))
 
 
+@router.post("/create-subscription-session")
+async def create_subscription_session(
+    _: User = Depends(current_admin_user),
+) -> SubscriptionSessionResponse:
+    try:
+        tenant_id = CURRENT_TENANT_ID_CONTEXTVAR.get()
+        session_id = fetch_stripe_checkout_session(tenant_id)
+        return SubscriptionSessionResponse(sessionId=session_id)
+
+    except Exception as e:
+        logger.exception("Failed to create resubscription session")
+        raise HTTPException(status_code=500, detail=str(e))
+
+
 @router.post("/impersonate")
 async def impersonate_user(
     impersonate_request: ImpersonateRequest,

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

@@ -6,6 +6,7 @@ import stripe
 from ee.onyx.configs.app_configs import STRIPE_PRICE_ID
 from ee.onyx.configs.app_configs import STRIPE_SECRET_KEY
 from ee.onyx.server.tenants.access import generate_data_plane_token
+from ee.onyx.server.tenants.models import BillingInformation
 from onyx.configs.app_configs import CONTROL_PLANE_API_BASE_URL
 from onyx.utils.logger import setup_logger
 
@@ -14,6 +15,19 @@ stripe.api_key = STRIPE_SECRET_KEY
 logger = setup_logger()
 
 
+def fetch_stripe_checkout_session(tenant_id: str) -> str:
+    token = generate_data_plane_token()
+    headers = {
+        "Authorization": f"Bearer {token}",
+        "Content-Type": "application/json",
+    }
+    url = f"{CONTROL_PLANE_API_BASE_URL}/create-checkout-session"
+    params = {"tenant_id": tenant_id}
+    response = requests.post(url, headers=headers, params=params)
+    response.raise_for_status()
+    return response.json()["sessionId"]
+
+
 def fetch_tenant_stripe_information(tenant_id: str) -> dict:
     token = generate_data_plane_token()
     headers = {
@@ -27,7 +41,7 @@ def fetch_tenant_stripe_information(tenant_id: str) -> dict:
     return response.json()
 
 
-def fetch_billing_information(tenant_id: str) -> dict:
+def fetch_billing_information(tenant_id: str) -> BillingInformation:
     logger.info("Fetching billing information")
     token = generate_data_plane_token()
     headers = {
@@ -38,7 +52,7 @@ def fetch_billing_information(tenant_id: str) -> dict:
     params = {"tenant_id": tenant_id}
     response = requests.get(url, headers=headers, params=params)
     response.raise_for_status()
-    billing_info = response.json()
+    billing_info = BillingInformation(**response.json())
     return billing_info
 
 

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

@@ -1,7 +1,8 @@
+from datetime import datetime
+
 from pydantic import BaseModel
 
-from onyx.configs.constants import NotificationType
-from onyx.server.settings.models import GatingType
+from onyx.server.settings.models import ApplicationStatus
 
 
 class CheckoutSessionCreationRequest(BaseModel):
@@ -15,15 +16,24 @@ class CreateTenantRequest(BaseModel):
 
 class ProductGatingRequest(BaseModel):
     tenant_id: str
-    product_gating: GatingType
-    notification: NotificationType | None = None
+    application_status: ApplicationStatus
+
+
+class SubscriptionStatusResponse(BaseModel):
+    subscribed: bool
 
 
 class BillingInformation(BaseModel):
+    stripe_subscription_id: str
+    status: str
+    current_period_start: datetime
+    current_period_end: datetime
+    number_of_seats: int
+    cancel_at_period_end: bool
+    canceled_at: datetime | None
+    trial_start: datetime | None
+    trial_end: datetime | None
     seats: int
-    subscription_status: str
-    billing_start: str
-    billing_end: str
     payment_method_enabled: bool
 
 
@@ -48,3 +58,12 @@ class TenantDeletionPayload(BaseModel):
 
 class AnonymousUserPath(BaseModel):
     anonymous_user_path: str | None
+
+
+class ProductGatingResponse(BaseModel):
+    updated: bool
+    error: str | None
+
+
+class SubscriptionSessionResponse(BaseModel):
+    sessionId: str

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

@@ -0,0 +1,51 @@
+from typing import cast
+
+from ee.onyx.configs.app_configs import GATED_TENANTS_KEY
+from onyx.configs.constants import ONYX_CLOUD_TENANT_ID
+from onyx.redis.redis_pool import get_redis_client
+from onyx.redis.redis_pool import get_redis_replica_client
+from onyx.server.settings.models import ApplicationStatus
+from onyx.server.settings.store import load_settings
+from onyx.server.settings.store import store_settings
+from onyx.setup import setup_logger
+from shared_configs.contextvars import CURRENT_TENANT_ID_CONTEXTVAR
+
+logger = setup_logger()
+
+
+def update_tenant_gating(tenant_id: str, status: ApplicationStatus) -> None:
+    redis_client = get_redis_client(tenant_id=ONYX_CLOUD_TENANT_ID)
+
+    # Store the full status
+    status_key = f"tenant:{tenant_id}:status"
+    redis_client.set(status_key, status.value)
+
+    # Maintain the GATED_ACCESS set
+    if status == ApplicationStatus.GATED_ACCESS:
+        redis_client.sadd(GATED_TENANTS_KEY, tenant_id)
+    else:
+        redis_client.srem(GATED_TENANTS_KEY, tenant_id)
+
+
+def store_product_gating(tenant_id: str, application_status: ApplicationStatus) -> None:
+    try:
+        token = CURRENT_TENANT_ID_CONTEXTVAR.set(tenant_id)
+
+        settings = load_settings()
+        settings.application_status = application_status
+        store_settings(settings)
+
+        # Store gated tenant information in Redis
+        update_tenant_gating(tenant_id, application_status)
+
+        if token is not None:
+            CURRENT_TENANT_ID_CONTEXTVAR.reset(token)
+
+    except Exception:
+        logger.exception("Failed to gate product")
+        raise
+
+
+def get_gated_tenants() -> set[str]:
+    redis_client = get_redis_replica_client(tenant_id=ONYX_CLOUD_TENANT_ID)
+    return cast(set[str], redis_client.smembers(GATED_TENANTS_KEY))

backend/model_server/constants.py

@@ -28,3 +28,9 @@ class EmbeddingModelTextType:
     @staticmethod
     def get_type(provider: EmbeddingProvider, text_type: EmbedTextType) -> str:
         return EmbeddingModelTextType.PROVIDER_TEXT_TYPE_MAP[provider][text_type]
+
+
+class GPUStatus:
+    CUDA = "cuda"
+    MAC_MPS = "mps"
+    NONE = "none"

backend/model_server/encoders.py

@@ -12,6 +12,7 @@ import voyageai  # type: ignore
 from cohere import AsyncClient as CohereAsyncClient
 from fastapi import APIRouter
 from fastapi import HTTPException
+from fastapi import Request
 from google.oauth2 import service_account  # type: ignore
 from litellm import aembedding
 from litellm.exceptions import RateLimitError
@@ -320,6 +321,7 @@ async def embed_text(
     prefix: str | None,
     api_url: str | None,
     api_version: str | None,
+    gpu_type: str = "UNKNOWN",
 ) -> list[Embedding]:
     if not all(texts):
         logger.error("Empty strings provided for embedding")
@@ -373,8 +375,11 @@ async def embed_text(
 
         elapsed = time.monotonic() - start
         logger.info(
-            f"Successfully embedded {len(texts)} texts with {total_chars} total characters "
-            f"with provider {provider_type} in {elapsed:.2f}"
+            f"event=embedding_provider "
+            f"texts={len(texts)} "
+            f"chars={total_chars} "
+            f"provider={provider_type} "
+            f"elapsed={elapsed:.2f}"
         )
     elif model_name is not None:
         logger.info(
@@ -403,6 +408,14 @@ async def embed_text(
             f"Successfully embedded {len(texts)} texts with {total_chars} total characters "
             f"with local model {model_name} in {elapsed:.2f}"
         )
+        logger.info(
+            f"event=embedding_model "
+            f"texts={len(texts)} "
+            f"chars={total_chars} "
+            f"model={model_name} "
+            f"gpu={gpu_type} "
+            f"elapsed={elapsed:.2f}"
+        )
     else:
         logger.error("Neither model name nor provider specified for embedding")
         raise ValueError(
@@ -455,8 +468,15 @@ async def litellm_rerank(
 
 
 @router.post("/bi-encoder-embed")
-async def process_embed_request(
+async def route_bi_encoder_embed(
+    request: Request,
     embed_request: EmbedRequest,
+) -> EmbedResponse:
+    return await process_embed_request(embed_request, request.app.state.gpu_type)
+
+
+async def process_embed_request(
+    embed_request: EmbedRequest, gpu_type: str = "UNKNOWN"
 ) -> EmbedResponse:
     if not embed_request.texts:
         raise HTTPException(status_code=400, detail="No texts to be embedded")
@@ -484,6 +504,7 @@ async def process_embed_request(
             api_url=embed_request.api_url,
             api_version=embed_request.api_version,
             prefix=prefix,
+            gpu_type=gpu_type,
         )
         return EmbedResponse(embeddings=embeddings)
     except RateLimitError as e:

backend/model_server/main.py

@@ -16,6 +16,7 @@ from model_server.custom_models import router as custom_models_router
 from model_server.custom_models import warm_up_intent_model
 from model_server.encoders import router as encoders_router
 from model_server.management_endpoints import router as management_router
+from model_server.utils import get_gpu_type
 from onyx import __version__
 from onyx.utils.logger import setup_logger
 from shared_configs.configs import INDEXING_ONLY
@@ -58,12 +59,10 @@ def _move_files_recursively(source: Path, dest: Path, overwrite: bool = False) -
 
 @asynccontextmanager
 async def lifespan(app: FastAPI) -> AsyncGenerator:
-    if torch.cuda.is_available():
-        logger.notice("CUDA GPU is available")
-    elif torch.backends.mps.is_available():
-        logger.notice("Mac MPS is available")
-    else:
-        logger.notice("GPU is not available, using CPU")
+    gpu_type = get_gpu_type()
+    logger.notice(f"Torch GPU Detection: gpu_type={gpu_type}")
+
+    app.state.gpu_type = gpu_type
 
     if TEMP_HF_CACHE_PATH.is_dir():
         logger.notice("Moving contents of temp_huggingface to huggingface cache.")

backend/model_server/management_endpoints.py

@@ -1,7 +1,9 @@
-import torch
 from fastapi import APIRouter
 from fastapi import Response
 
+from model_server.constants import GPUStatus
+from model_server.utils import get_gpu_type
+
 router = APIRouter(prefix="/api")
 
 
@@ -11,10 +13,7 @@ async def healthcheck() -> Response:
 
 
 @router.get("/gpu-status")
-async def gpu_status() -> dict[str, bool | str]:
-    if torch.cuda.is_available():
-        return {"gpu_available": True, "type": "cuda"}
-    elif torch.backends.mps.is_available():
-        return {"gpu_available": True, "type": "mps"}
-    else:
-        return {"gpu_available": False, "type": "none"}
+async def route_gpu_status() -> dict[str, bool | str]:
+    gpu_type = get_gpu_type()
+    gpu_available = gpu_type != GPUStatus.NONE
+    return {"gpu_available": gpu_available, "type": gpu_type}

backend/model_server/utils.py

@@ -8,6 +8,9 @@ from typing import Any
 from typing import cast
 from typing import TypeVar
 
+import torch
+
+from model_server.constants import GPUStatus
 from onyx.utils.logger import setup_logger
 
 logger = setup_logger()
@@ -58,3 +61,12 @@ def simple_log_function_time(
             return cast(F, wrapped_sync_func)
 
     return decorator
+
+
+def get_gpu_type() -> str:
+    if torch.cuda.is_available():
+        return GPUStatus.CUDA
+    if torch.backends.mps.is_available():
+        return GPUStatus.MAC_MPS
+
+    return GPUStatus.NONE

backend/onyx/agents/agent_search/basic/graph_builder.py

@@ -85,7 +85,7 @@ if __name__ == "__main__":
 
     graph = basic_graph_builder()
     compiled_graph = graph.compile()
-    input = BasicInput(_unused=True)
+    input = BasicInput(unused=True)
     primary_llm, fast_llm = get_default_llms()
     with get_session_context_manager() as db_session:
         config, _ = get_test_config(

backend/onyx/agents/agent_search/basic/states.py

@@ -17,7 +17,7 @@ from onyx.agents.agent_search.orchestration.states import ToolChoiceUpdate
 class BasicInput(BaseModel):
     # Langgraph needs a nonempty input, but we pass in all static
     # data through a RunnableConfig.
-    _unused: bool = True
+    unused: bool = True
 
 
 ## Graph Output State

backend/onyx/agents/agent_search/core_state.py

@@ -9,7 +9,6 @@ class CoreState(BaseModel):
     This is the core state that is shared across all subgraphs.
     """
 
-    base_question: str = ""
     log_messages: Annotated[list[str], add] = []
 
 
@@ -18,4 +17,4 @@ class SubgraphCoreState(BaseModel):
     This is the core state that is shared across all subgraphs.
     """
 
-    log_messages: Annotated[list[str], add]
+    log_messages: Annotated[list[str], add] = []

backend/onyx/agents/agent_search/deep_search/initial/generate_individual_sub_answer/nodes/check_sub_answer.py

@@ -1,8 +1,8 @@
 from datetime import datetime
 from typing import cast
 
+from langchain_core.messages import BaseMessage
 from langchain_core.messages import HumanMessage
-from langchain_core.messages import merge_message_runs
 from langchain_core.runnables.config import RunnableConfig
 
 from onyx.agents.agent_search.deep_search.initial.generate_individual_sub_answer.states import (
@@ -12,14 +12,43 @@ from onyx.agents.agent_search.deep_search.initial.generate_individual_sub_answer
     SubQuestionAnswerCheckUpdate,
 )
 from onyx.agents.agent_search.models import GraphConfig
+from onyx.agents.agent_search.shared_graph_utils.agent_prompt_ops import (
+    binary_string_test,
+)
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AGENT_LLM_RATELIMIT_MESSAGE,
+)
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AGENT_LLM_TIMEOUT_MESSAGE,
+)
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AGENT_POSITIVE_VALUE_STR,
+)
+from onyx.agents.agent_search.shared_graph_utils.constants import AgentLLMErrorType
+from onyx.agents.agent_search.shared_graph_utils.models import AgentErrorLog
+from onyx.agents.agent_search.shared_graph_utils.models import LLMNodeErrorStrings
 from onyx.agents.agent_search.shared_graph_utils.utils import (
     get_langgraph_node_log_string,
 )
 from onyx.agents.agent_search.shared_graph_utils.utils import parse_question_id
+from onyx.configs.agent_configs import AGENT_TIMEOUT_OVERRIDE_LLM_SUBANSWER_CHECK
+from onyx.llm.chat_llm import LLMRateLimitError
+from onyx.llm.chat_llm import LLMTimeoutError
 from onyx.prompts.agent_search import SUB_ANSWER_CHECK_PROMPT
 from onyx.prompts.agent_search import UNKNOWN_ANSWER
+from onyx.utils.logger import setup_logger
+from onyx.utils.timing import log_function_time
+
+logger = setup_logger()
+
+_llm_node_error_strings = LLMNodeErrorStrings(
+    timeout="LLM Timeout Error. The sub-answer will be treated as 'relevant'",
+    rate_limit="LLM Rate Limit Error. The sub-answer will be treated as 'relevant'",
+    general_error="General LLM Error. The sub-answer will be treated as 'relevant'",
+)
 
 
+@log_function_time(print_only=True)
 def check_sub_answer(
     state: AnswerQuestionState, config: RunnableConfig
 ) -> SubQuestionAnswerCheckUpdate:
@@ -53,14 +82,40 @@ def check_sub_answer(
 
     graph_config = cast(GraphConfig, config["metadata"]["config"])
     fast_llm = graph_config.tooling.fast_llm
-    response = list(
-        fast_llm.stream(
+    agent_error: AgentErrorLog | None = None
+    response: BaseMessage | None = None
+    try:
+        response = fast_llm.invoke(
             prompt=msg,
+            timeout_override=AGENT_TIMEOUT_OVERRIDE_LLM_SUBANSWER_CHECK,
         )
-    )
 
-    quality_str: str = merge_message_runs(response, chunk_separator="")[0].content
-    answer_quality = "yes" in quality_str.lower()
+        quality_str: str = cast(str, response.content)
+        answer_quality = binary_string_test(
+            text=quality_str, positive_value=AGENT_POSITIVE_VALUE_STR
+        )
+        log_result = f"Answer quality: {quality_str}"
+
+    except LLMTimeoutError:
+        agent_error = AgentErrorLog(
+            error_type=AgentLLMErrorType.TIMEOUT,
+            error_message=AGENT_LLM_TIMEOUT_MESSAGE,
+            error_result=_llm_node_error_strings.timeout,
+        )
+        answer_quality = True
+        log_result = agent_error.error_result
+        logger.error("LLM Timeout Error - check sub answer")
+
+    except LLMRateLimitError:
+        agent_error = AgentErrorLog(
+            error_type=AgentLLMErrorType.RATE_LIMIT,
+            error_message=AGENT_LLM_RATELIMIT_MESSAGE,
+            error_result=_llm_node_error_strings.rate_limit,
+        )
+
+        answer_quality = True
+        log_result = agent_error.error_result
+        logger.error("LLM Rate Limit Error - check sub answer")
 
     return SubQuestionAnswerCheckUpdate(
         answer_quality=answer_quality,
@@ -69,7 +124,7 @@ def check_sub_answer(
                 graph_component="initial  - generate individual sub answer",
                 node_name="check sub answer",
                 node_start_time=node_start_time,
-                result=f"Answer quality: {quality_str}",
+                result=log_result,
             )
         ],
     )

backend/onyx/agents/agent_search/deep_search/initial/generate_individual_sub_answer/nodes/generate_sub_answer.py

@@ -16,6 +16,23 @@ from onyx.agents.agent_search.models import GraphConfig
 from onyx.agents.agent_search.shared_graph_utils.agent_prompt_ops import (
     build_sub_question_answer_prompt,
 )
+from onyx.agents.agent_search.shared_graph_utils.calculations import (
+    dedup_sort_inference_section_list,
+)
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AGENT_LLM_RATELIMIT_MESSAGE,
+)
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AGENT_LLM_TIMEOUT_MESSAGE,
+)
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AgentLLMErrorType,
+)
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    LLM_ANSWER_ERROR_MESSAGE,
+)
+from onyx.agents.agent_search.shared_graph_utils.models import AgentErrorLog
+from onyx.agents.agent_search.shared_graph_utils.models import LLMNodeErrorStrings
 from onyx.agents.agent_search.shared_graph_utils.utils import get_answer_citation_ids
 from onyx.agents.agent_search.shared_graph_utils.utils import (
     get_langgraph_node_log_string,
@@ -30,12 +47,23 @@ from onyx.chat.models import StreamStopInfo
 from onyx.chat.models import StreamStopReason
 from onyx.chat.models import StreamType
 from onyx.configs.agent_configs import AGENT_MAX_ANSWER_CONTEXT_DOCS
+from onyx.configs.agent_configs import AGENT_TIMEOUT_OVERRIDE_LLM_SUBANSWER_GENERATION
+from onyx.llm.chat_llm import LLMRateLimitError
+from onyx.llm.chat_llm import LLMTimeoutError
 from onyx.prompts.agent_search import NO_RECOVERED_DOCS
 from onyx.utils.logger import setup_logger
+from onyx.utils.timing import log_function_time
 
 logger = setup_logger()
 
+_llm_node_error_strings = LLMNodeErrorStrings(
+    timeout="LLM Timeout Error. A sub-answer could not be constructed and the sub-question will be ignored.",
+    rate_limit="LLM Rate Limit Error. A sub-answer could not be constructed and the sub-question will be ignored.",
+    general_error="General LLM Error. A sub-answer could not be constructed and the sub-question will be ignored.",
+)
 
+
+@log_function_time(print_only=True)
 def generate_sub_answer(
     state: AnswerQuestionState,
     config: RunnableConfig,
@@ -51,12 +79,17 @@ def generate_sub_answer(
     state.verified_reranked_documents
     level, question_num = parse_question_id(state.question_id)
     context_docs = state.context_documents[:AGENT_MAX_ANSWER_CONTEXT_DOCS]
+
+    context_docs = dedup_sort_inference_section_list(context_docs)
+
     persona_contextualized_prompt = get_persona_agent_prompt_expressions(
         graph_config.inputs.search_request.persona
     ).contextualized_prompt
 
     if len(context_docs) == 0:
         answer_str = NO_RECOVERED_DOCS
+        cited_documents: list = []
+        log_results = "No documents retrieved"
         write_custom_event(
             "sub_answers",
             AgentAnswerPiece(
@@ -79,41 +112,67 @@ def generate_sub_answer(
 
         response: list[str | list[str | dict[str, Any]]] = []
         dispatch_timings: list[float] = []
-        for message in fast_llm.stream(
-            prompt=msg,
-        ):
-            # TODO: in principle, the answer here COULD contain images, but we don't support that yet
-            content = message.content
-            if not isinstance(content, str):
-                raise ValueError(
-                    f"Expected content to be a string, but got {type(content)}"
+
+        agent_error: AgentErrorLog | None = None
+
+        try:
+            for message in fast_llm.stream(
+                prompt=msg,
+                timeout_override=AGENT_TIMEOUT_OVERRIDE_LLM_SUBANSWER_GENERATION,
+            ):
+                # TODO: in principle, the answer here COULD contain images, but we don't support that yet
+                content = message.content
+                if not isinstance(content, str):
+                    raise ValueError(
+                        f"Expected content to be a string, but got {type(content)}"
+                    )
+                start_stream_token = datetime.now()
+                write_custom_event(
+                    "sub_answers",
+                    AgentAnswerPiece(
+                        answer_piece=content,
+                        level=level,
+                        level_question_num=question_num,
+                        answer_type="agent_sub_answer",
+                    ),
+                    writer,
                 )
-            start_stream_token = datetime.now()
-            write_custom_event(
-                "sub_answers",
-                AgentAnswerPiece(
-                    answer_piece=content,
-                    level=level,
-                    level_question_num=question_num,
-                    answer_type="agent_sub_answer",
-                ),
-                writer,
-            )
-            end_stream_token = datetime.now()
-            dispatch_timings.append(
-                (end_stream_token - start_stream_token).microseconds
-            )
-            response.append(content)
+                end_stream_token = datetime.now()
+                dispatch_timings.append(
+                    (end_stream_token - start_stream_token).microseconds
+                )
+                response.append(content)
 
-        answer_str = merge_message_runs(response, chunk_separator="")[0].content
-        logger.debug(
-            f"Average dispatch time: {sum(dispatch_timings) / len(dispatch_timings)}"
-        )
+        except LLMTimeoutError:
+            agent_error = AgentErrorLog(
+                error_type=AgentLLMErrorType.TIMEOUT,
+                error_message=AGENT_LLM_TIMEOUT_MESSAGE,
+                error_result=_llm_node_error_strings.timeout,
+            )
+            logger.error("LLM Timeout Error - generate sub answer")
+        except LLMRateLimitError:
+            agent_error = AgentErrorLog(
+                error_type=AgentLLMErrorType.RATE_LIMIT,
+                error_message=AGENT_LLM_RATELIMIT_MESSAGE,
+                error_result=_llm_node_error_strings.rate_limit,
+            )
+            logger.error("LLM Rate Limit Error - generate sub answer")
 
-    answer_citation_ids = get_answer_citation_ids(answer_str)
-    cited_documents = [
-        context_docs[id] for id in answer_citation_ids if id < len(context_docs)
-    ]
+        if agent_error:
+            answer_str = LLM_ANSWER_ERROR_MESSAGE
+            cited_documents = []
+            log_results = (
+                agent_error.error_result
+                or "Sub-answer generation failed due to LLM error"
+            )
+
+        else:
+            answer_str = merge_message_runs(response, chunk_separator="")[0].content
+            answer_citation_ids = get_answer_citation_ids(answer_str)
+            cited_documents = [
+                context_docs[id] for id in answer_citation_ids if id < len(context_docs)
+            ]
+            log_results = None
 
     stop_event = StreamStopInfo(
         stop_reason=StreamStopReason.FINISHED,
@@ -131,7 +190,7 @@ def generate_sub_answer(
                 graph_component="initial - generate individual sub answer",
                 node_name="generate sub answer",
                 node_start_time=node_start_time,
-                result="",
+                result=log_results or "",
             )
         ],
     )

backend/onyx/agents/agent_search/deep_search/initial/generate_individual_sub_answer/states.py

@@ -42,10 +42,8 @@ class SubQuestionRetrievalIngestionUpdate(LoggerUpdate, BaseModel):
 
 
 class SubQuestionAnsweringInput(SubgraphCoreState):
-    question: str = ""
-    question_id: str = (
-        ""  # 0_0 is original question, everything else is <level>_<question_num>.
-    )
+    question: str
+    question_id: str
     # level 0 is original question and first decomposition, level 1 is follow up, etc
     # question_num is a unique number per original question per level.
 

backend/onyx/agents/agent_search/deep_search/initial/generate_initial_answer/nodes/generate_initial_answer.py

@@ -26,14 +26,31 @@ from onyx.agents.agent_search.shared_graph_utils.agent_prompt_ops import (
 from onyx.agents.agent_search.shared_graph_utils.agent_prompt_ops import (
     trim_prompt_piece,
 )
+from onyx.agents.agent_search.shared_graph_utils.calculations import (
+    get_answer_generation_documents,
+)
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AGENT_LLM_RATELIMIT_MESSAGE,
+)
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AGENT_LLM_TIMEOUT_MESSAGE,
+)
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AgentLLMErrorType,
+)
+from onyx.agents.agent_search.shared_graph_utils.models import AgentErrorLog
 from onyx.agents.agent_search.shared_graph_utils.models import InitialAgentResultStats
+from onyx.agents.agent_search.shared_graph_utils.models import LLMNodeErrorStrings
 from onyx.agents.agent_search.shared_graph_utils.operators import (
-    dedup_inference_sections,
+    dedup_inference_section_list,
 )
 from onyx.agents.agent_search.shared_graph_utils.utils import (
     dispatch_main_answer_stop_info,
 )
 from onyx.agents.agent_search.shared_graph_utils.utils import format_docs
+from onyx.agents.agent_search.shared_graph_utils.utils import (
+    get_deduplicated_structured_subquestion_documents,
+)
 from onyx.agents.agent_search.shared_graph_utils.utils import (
     get_langgraph_node_log_string,
 )
@@ -42,12 +59,16 @@ from onyx.agents.agent_search.shared_graph_utils.utils import remove_document_ci
 from onyx.agents.agent_search.shared_graph_utils.utils import write_custom_event
 from onyx.chat.models import AgentAnswerPiece
 from onyx.chat.models import ExtendedToolResponse
+from onyx.chat.models import StreamingError
 from onyx.configs.agent_configs import AGENT_MAX_ANSWER_CONTEXT_DOCS
+from onyx.configs.agent_configs import AGENT_MAX_STREAMED_DOCS_FOR_INITIAL_ANSWER
 from onyx.configs.agent_configs import AGENT_MIN_ORIG_QUESTION_DOCS
-from onyx.context.search.models import InferenceSection
-from onyx.prompts.agent_search import (
-    INITIAL_ANSWER_PROMPT_W_SUB_QUESTIONS,
+from onyx.configs.agent_configs import (
+    AGENT_TIMEOUT_OVERRIDE_LLM_INITIAL_ANSWER_GENERATION,
 )
+from onyx.llm.chat_llm import LLMRateLimitError
+from onyx.llm.chat_llm import LLMTimeoutError
+from onyx.prompts.agent_search import INITIAL_ANSWER_PROMPT_W_SUB_QUESTIONS
 from onyx.prompts.agent_search import (
     INITIAL_ANSWER_PROMPT_WO_SUB_QUESTIONS,
 )
@@ -56,8 +77,16 @@ from onyx.prompts.agent_search import (
 )
 from onyx.prompts.agent_search import UNKNOWN_ANSWER
 from onyx.tools.tool_implementations.search.search_tool import yield_search_responses
+from onyx.utils.timing import log_function_time
+
+_llm_node_error_strings = LLMNodeErrorStrings(
+    timeout="LLM Timeout Error. The initial answer could not be generated.",
+    rate_limit="LLM Rate Limit Error. The initial answer could not be generated.",
+    general_error="General LLM Error. The initial answer could not be generated.",
+)
 
 
+@log_function_time(print_only=True)
 def generate_initial_answer(
     state: SubQuestionRetrievalState,
     config: RunnableConfig,
@@ -73,15 +102,19 @@ def generate_initial_answer(
     question = graph_config.inputs.search_request.query
     prompt_enrichment_components = get_prompt_enrichment_components(graph_config)
 
-    sub_questions_cited_documents = state.cited_documents
+    # get all documents cited in sub-questions
+    structured_subquestion_docs = get_deduplicated_structured_subquestion_documents(
+        state.sub_question_results
+    )
+
     orig_question_retrieval_documents = state.orig_question_retrieved_documents
 
-    consolidated_context_docs: list[InferenceSection] = sub_questions_cited_documents
+    consolidated_context_docs = structured_subquestion_docs.cited_documents
     counter = 0
     for original_doc_number, original_doc in enumerate(
         orig_question_retrieval_documents
     ):
-        if original_doc_number not in sub_questions_cited_documents:
+        if original_doc_number not in structured_subquestion_docs.cited_documents:
             if (
                 counter <= AGENT_MIN_ORIG_QUESTION_DOCS
                 or len(consolidated_context_docs) < AGENT_MAX_ANSWER_CONTEXT_DOCS
@@ -90,15 +123,18 @@ def generate_initial_answer(
                 counter += 1
 
     # sort docs by their scores - though the scores refer to different questions
-    relevant_docs = dedup_inference_sections(
-        consolidated_context_docs, consolidated_context_docs
-    )
+    relevant_docs = dedup_inference_section_list(consolidated_context_docs)
 
     sub_questions: list[str] = []
-    streamed_documents = (
-        relevant_docs
-        if len(relevant_docs) > 0
-        else state.orig_question_retrieved_documents[:15]
+
+    # Create the list of documents to stream out. Start with the
+    # ones that wil be in the context (or, if len == 0, use docs
+    # that were retrieved for the original question)
+    answer_generation_documents = get_answer_generation_documents(
+        relevant_docs=relevant_docs,
+        context_documents=structured_subquestion_docs.context_documents,
+        original_question_docs=orig_question_retrieval_documents,
+        max_docs=AGENT_MAX_STREAMED_DOCS_FOR_INITIAL_ANSWER,
     )
 
     # Use the query info from the base document retrieval
@@ -108,11 +144,13 @@ def generate_initial_answer(
         graph_config.tooling.search_tool
     ), "search_tool must be provided for agentic search"
 
-    relevance_list = relevance_from_docs(relevant_docs)
+    relevance_list = relevance_from_docs(
+        answer_generation_documents.streaming_documents
+    )
     for tool_response in yield_search_responses(
         query=question,
-        reranked_sections=streamed_documents,
-        final_context_sections=streamed_documents,
+        reranked_sections=answer_generation_documents.streaming_documents,
+        final_context_sections=answer_generation_documents.context_documents,
         search_query_info=query_info,
         get_section_relevance=lambda: relevance_list,
         search_tool=graph_config.tooling.search_tool,
@@ -128,7 +166,7 @@ def generate_initial_answer(
             writer,
         )
 
-    if len(relevant_docs) == 0:
+    if len(answer_generation_documents.context_documents) == 0:
         write_custom_event(
             "initial_agent_answer",
             AgentAnswerPiece(
@@ -194,7 +232,7 @@ def generate_initial_answer(
 
         model = graph_config.tooling.fast_llm
 
-        doc_context = format_docs(relevant_docs)
+        doc_context = format_docs(answer_generation_documents.context_documents)
         doc_context = trim_prompt_piece(
             config=model.config,
             prompt_piece=doc_context,
@@ -224,30 +262,82 @@ def generate_initial_answer(
 
         streamed_tokens: list[str | list[str | dict[str, Any]]] = [""]
         dispatch_timings: list[float] = []
-        for message in model.stream(msg):
-            # TODO: in principle, the answer here COULD contain images, but we don't support that yet
-            content = message.content
-            if not isinstance(content, str):
-                raise ValueError(
-                    f"Expected content to be a string, but got {type(content)}"
-                )
-            start_stream_token = datetime.now()
 
+        agent_error: AgentErrorLog | None = None
+
+        try:
+            for message in model.stream(
+                msg,
+                timeout_override=AGENT_TIMEOUT_OVERRIDE_LLM_INITIAL_ANSWER_GENERATION,
+            ):
+                # TODO: in principle, the answer here COULD contain images, but we don't support that yet
+                content = message.content
+                if not isinstance(content, str):
+                    raise ValueError(
+                        f"Expected content to be a string, but got {type(content)}"
+                    )
+                start_stream_token = datetime.now()
+
+                write_custom_event(
+                    "initial_agent_answer",
+                    AgentAnswerPiece(
+                        answer_piece=content,
+                        level=0,
+                        level_question_num=0,
+                        answer_type="agent_level_answer",
+                    ),
+                    writer,
+                )
+                end_stream_token = datetime.now()
+                dispatch_timings.append(
+                    (end_stream_token - start_stream_token).microseconds
+                )
+                streamed_tokens.append(content)
+
+        except LLMTimeoutError:
+            agent_error = AgentErrorLog(
+                error_type=AgentLLMErrorType.TIMEOUT,
+                error_message=AGENT_LLM_TIMEOUT_MESSAGE,
+                error_result=_llm_node_error_strings.timeout,
+            )
+            logger.error("LLM Timeout Error - generate initial answer")
+
+        except LLMRateLimitError:
+            agent_error = AgentErrorLog(
+                error_type=AgentLLMErrorType.RATE_LIMIT,
+                error_message=AGENT_LLM_RATELIMIT_MESSAGE,
+                error_result=_llm_node_error_strings.rate_limit,
+            )
+            logger.error("LLM Rate Limit Error - generate initial answer")
+
+        if agent_error:
             write_custom_event(
                 "initial_agent_answer",
-                AgentAnswerPiece(
-                    answer_piece=content,
-                    level=0,
-                    level_question_num=0,
-                    answer_type="agent_level_answer",
+                StreamingError(
+                    error=AGENT_LLM_TIMEOUT_MESSAGE,
                 ),
                 writer,
             )
-            end_stream_token = datetime.now()
-            dispatch_timings.append(
-                (end_stream_token - start_stream_token).microseconds
+            return InitialAnswerUpdate(
+                initial_answer=None,
+                answer_error=AgentErrorLog(
+                    error_message=agent_error.error_message or "An LLM error occurred",
+                    error_type=agent_error.error_type,
+                    error_result=agent_error.error_result,
+                ),
+                initial_agent_stats=None,
+                generated_sub_questions=sub_questions,
+                agent_base_end_time=None,
+                agent_base_metrics=None,
+                log_messages=[
+                    get_langgraph_node_log_string(
+                        graph_component="initial - generate initial answer",
+                        node_name="generate initial answer",
+                        node_start_time=node_start_time,
+                        result=agent_error.error_result or "An LLM error occurred",
+                    )
+                ],
             )
-            streamed_tokens.append(content)
 
         logger.debug(
             f"Average dispatch time for initial answer: {sum(dispatch_timings) / len(dispatch_timings)}"

backend/onyx/agents/agent_search/deep_search/initial/generate_initial_answer/nodes/validate_initial_answer.py

@@ -10,8 +10,10 @@ from onyx.agents.agent_search.deep_search.main.states import (
 from onyx.agents.agent_search.shared_graph_utils.utils import (
     get_langgraph_node_log_string,
 )
+from onyx.utils.timing import log_function_time
 
 
+@log_function_time(print_only=True)
 def validate_initial_answer(
     state: SubQuestionRetrievalState,
 ) -> InitialAnswerQualityUpdate:
@@ -25,7 +27,7 @@ def validate_initial_answer(
         f"--------{node_start_time}--------Checking for base answer validity - for not set True/False manually"
     )
 
-    verdict = True
+    verdict = True  # not actually required as already streamed out. Refinement will do similar
 
     return InitialAnswerQualityUpdate(
         initial_answer_quality_eval=verdict,

backend/onyx/agents/agent_search/deep_search/initial/generate_sub_answers/nodes/decompose_orig_question.py

@@ -23,6 +23,8 @@ from onyx.agents.agent_search.models import GraphConfig
 from onyx.agents.agent_search.shared_graph_utils.agent_prompt_ops import (
     build_history_prompt,
 )
+from onyx.agents.agent_search.shared_graph_utils.models import BaseMessage_Content
+from onyx.agents.agent_search.shared_graph_utils.models import LLMNodeErrorStrings
 from onyx.agents.agent_search.shared_graph_utils.utils import dispatch_separated
 from onyx.agents.agent_search.shared_graph_utils.utils import (
     get_langgraph_node_log_string,
@@ -33,17 +35,30 @@ from onyx.chat.models import StreamStopReason
 from onyx.chat.models import StreamType
 from onyx.chat.models import SubQuestionPiece
 from onyx.configs.agent_configs import AGENT_NUM_DOCS_FOR_DECOMPOSITION
+from onyx.configs.agent_configs import (
+    AGENT_TIMEOUT_OVERRIDE_LLM_SUBQUESTION_GENERATION,
+)
+from onyx.llm.chat_llm import LLMRateLimitError
+from onyx.llm.chat_llm import LLMTimeoutError
 from onyx.prompts.agent_search import (
-    INITIAL_DECOMPOSITION_PROMPT_QUESTIONS_AFTER_SEARCH,
+    INITIAL_DECOMPOSITION_PROMPT_QUESTIONS_AFTER_SEARCH_ASSUMING_REFINEMENT,
 )
 from onyx.prompts.agent_search import (
-    INITIAL_QUESTION_DECOMPOSITION_PROMPT,
+    INITIAL_QUESTION_DECOMPOSITION_PROMPT_ASSUMING_REFINEMENT,
 )
 from onyx.utils.logger import setup_logger
+from onyx.utils.timing import log_function_time
 
 logger = setup_logger()
 
+_llm_node_error_strings = LLMNodeErrorStrings(
+    timeout="LLM Timeout Error. Sub-questions could not be generated.",
+    rate_limit="LLM Rate Limit Error. Sub-questions could not be generated.",
+    general_error="General LLM Error. Sub-questions could not be generated.",
+)
 
+
+@log_function_time(print_only=True)
 def decompose_orig_question(
     state: SubQuestionRetrievalState,
     config: RunnableConfig,
@@ -85,15 +100,15 @@ def decompose_orig_question(
             ]
         )
 
-        decomposition_prompt = (
-            INITIAL_DECOMPOSITION_PROMPT_QUESTIONS_AFTER_SEARCH.format(
-                question=question, sample_doc_str=sample_doc_str, history=history
-            )
+        decomposition_prompt = INITIAL_DECOMPOSITION_PROMPT_QUESTIONS_AFTER_SEARCH_ASSUMING_REFINEMENT.format(
+            question=question, sample_doc_str=sample_doc_str, history=history
         )
 
     else:
-        decomposition_prompt = INITIAL_QUESTION_DECOMPOSITION_PROMPT.format(
-            question=question, history=history
+        decomposition_prompt = (
+            INITIAL_QUESTION_DECOMPOSITION_PROMPT_ASSUMING_REFINEMENT.format(
+                question=question, history=history
+            )
         )
 
     # Start decomposition
@@ -112,32 +127,42 @@ def decompose_orig_question(
     )
 
     # dispatches custom events for subquestion tokens, adding in subquestion ids.
-    streamed_tokens = dispatch_separated(
-        model.stream(msg),
-        dispatch_subquestion(0, writer),
-        sep_callback=dispatch_subquestion_sep(0, writer),
-    )
 
-    stop_event = StreamStopInfo(
-        stop_reason=StreamStopReason.FINISHED,
-        stream_type=StreamType.SUB_QUESTIONS,
-        level=0,
-    )
-    write_custom_event("stream_finished", stop_event, writer)
+    streamed_tokens: list[BaseMessage_Content] = []
 
-    deomposition_response = merge_content(*streamed_tokens)
+    try:
+        streamed_tokens = dispatch_separated(
+            model.stream(
+                msg,
+                timeout_override=AGENT_TIMEOUT_OVERRIDE_LLM_SUBQUESTION_GENERATION,
+            ),
+            dispatch_subquestion(0, writer),
+            sep_callback=dispatch_subquestion_sep(0, writer),
+        )
 
-    # this call should only return strings. Commenting out for efficiency
-    # assert [type(tok) == str for tok in streamed_tokens]
+        decomposition_response = merge_content(*streamed_tokens)
 
-    # use no-op cast() instead of str() which runs code
-    # list_of_subquestions = clean_and_parse_list_string(cast(str, response))
-    list_of_subqs = cast(str, deomposition_response).split("\n")
+        list_of_subqs = cast(str, decomposition_response).split("\n")
 
-    decomp_list: list[str] = [sq.strip() for sq in list_of_subqs if sq.strip() != ""]
+        initial_sub_questions = [sq.strip() for sq in list_of_subqs if sq.strip() != ""]
+        log_result = f"decomposed original question into {len(initial_sub_questions)} subquestions"
+
+        stop_event = StreamStopInfo(
+            stop_reason=StreamStopReason.FINISHED,
+            stream_type=StreamType.SUB_QUESTIONS,
+            level=0,
+        )
+        write_custom_event("stream_finished", stop_event, writer)
+
+    except LLMTimeoutError as e:
+        logger.error("LLM Timeout Error - decompose orig question")
+        raise e  # fail loudly on this critical step
+    except LLMRateLimitError as e:
+        logger.error("LLM Rate Limit Error - decompose orig question")
+        raise e
 
     return InitialQuestionDecompositionUpdate(
-        initial_sub_questions=decomp_list,
+        initial_sub_questions=initial_sub_questions,
         agent_start_time=agent_start_time,
         agent_refined_start_time=None,
         agent_refined_end_time=None,
@@ -151,7 +176,7 @@ def decompose_orig_question(
                 graph_component="initial - generate sub answers",
                 node_name="decompose original question",
                 node_start_time=node_start_time,
-                result=f"decomposed original question into {len(decomp_list)} subquestions",
+                result=log_result,
             )
         ],
     )

backend/onyx/agents/agent_search/deep_search/main/graph_builder.py

@@ -26,8 +26,8 @@ from onyx.agents.agent_search.deep_search.main.nodes.decide_refinement_need impo
 from onyx.agents.agent_search.deep_search.main.nodes.extract_entities_terms import (
     extract_entities_terms,
 )
-from onyx.agents.agent_search.deep_search.main.nodes.generate_refined_answer import (
-    generate_refined_answer,
+from onyx.agents.agent_search.deep_search.main.nodes.generate_validate_refined_answer import (
+    generate_validate_refined_answer,
 )
 from onyx.agents.agent_search.deep_search.main.nodes.ingest_refined_sub_answers import (
     ingest_refined_sub_answers,
@@ -126,8 +126,8 @@ def main_graph_builder(test_mode: bool = False) -> StateGraph:
 
     # Node to generate the refined answer
     graph.add_node(
-        node="generate_refined_answer",
-        action=generate_refined_answer,
+        node="generate_validate_refined_answer",
+        action=generate_validate_refined_answer,
     )
 
     # Early node to extract the entities and terms from the initial answer,
@@ -215,11 +215,11 @@ def main_graph_builder(test_mode: bool = False) -> StateGraph:
 
     graph.add_edge(
         start_key="ingest_refined_sub_answers",
-        end_key="generate_refined_answer",
+        end_key="generate_validate_refined_answer",
     )
 
     graph.add_edge(
-        start_key="generate_refined_answer",
+        start_key="generate_validate_refined_answer",
         end_key="compare_answers",
     )
     graph.add_edge(
@@ -252,9 +252,7 @@ if __name__ == "__main__":
             db_session, primary_llm, fast_llm, search_request
         )
 
-        inputs = MainInput(
-            base_question=graph_config.inputs.search_request.query, log_messages=[]
-        )
+        inputs = MainInput(log_messages=[])
 
         for thing in compiled_graph.stream(
             input=inputs,

backend/onyx/agents/agent_search/deep_search/main/nodes/compare_answers.py

@@ -1,6 +1,7 @@
 from datetime import datetime
 from typing import cast
 
+from langchain_core.messages import BaseMessage
 from langchain_core.messages import HumanMessage
 from langchain_core.runnables import RunnableConfig
 from langgraph.types import StreamWriter
@@ -10,16 +11,51 @@ from onyx.agents.agent_search.deep_search.main.states import (
 )
 from onyx.agents.agent_search.deep_search.main.states import MainState
 from onyx.agents.agent_search.models import GraphConfig
+from onyx.agents.agent_search.shared_graph_utils.agent_prompt_ops import (
+    binary_string_test,
+)
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AGENT_LLM_RATELIMIT_MESSAGE,
+)
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AGENT_LLM_TIMEOUT_MESSAGE,
+)
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AGENT_POSITIVE_VALUE_STR,
+)
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AgentLLMErrorType,
+)
+from onyx.agents.agent_search.shared_graph_utils.models import AgentErrorLog
+from onyx.agents.agent_search.shared_graph_utils.models import LLMNodeErrorStrings
 from onyx.agents.agent_search.shared_graph_utils.utils import (
     get_langgraph_node_log_string,
 )
 from onyx.agents.agent_search.shared_graph_utils.utils import write_custom_event
 from onyx.chat.models import RefinedAnswerImprovement
+from onyx.configs.agent_configs import AGENT_TIMEOUT_OVERRIDE_LLM_COMPARE_ANSWERS
+from onyx.llm.chat_llm import LLMRateLimitError
+from onyx.llm.chat_llm import LLMTimeoutError
 from onyx.prompts.agent_search import (
     INITIAL_REFINED_ANSWER_COMPARISON_PROMPT,
 )
+from onyx.utils.logger import setup_logger
+from onyx.utils.timing import log_function_time
+
+logger = setup_logger()
+
+_llm_node_error_strings = LLMNodeErrorStrings(
+    timeout="The LLM timed out, and the answers could not be compared.",
+    rate_limit="The LLM encountered a rate limit, and the answers could not be compared.",
+    general_error="The LLM encountered an error, and the answers could not be compared.",
+)
+
+_ANSWER_QUALITY_NOT_SUFFICIENT_MESSAGE = (
+    "Answer quality is not sufficient, so stay with the initial answer."
+)
 
 
+@log_function_time(print_only=True)
 def compare_answers(
     state: MainState, config: RunnableConfig, writer: StreamWriter = lambda _: None
 ) -> InitialRefinedAnswerComparisonUpdate:
@@ -34,21 +70,75 @@ def compare_answers(
     initial_answer = state.initial_answer
     refined_answer = state.refined_answer
 
+    # if answer quality is not sufficient, then stay with the initial answer
+    if not state.refined_answer_quality:
+        write_custom_event(
+            "refined_answer_improvement",
+            RefinedAnswerImprovement(
+                refined_answer_improvement=False,
+            ),
+            writer,
+        )
+
+        return InitialRefinedAnswerComparisonUpdate(
+            refined_answer_improvement_eval=False,
+            log_messages=[
+                get_langgraph_node_log_string(
+                    graph_component="main",
+                    node_name="compare answers",
+                    node_start_time=node_start_time,
+                    result=_ANSWER_QUALITY_NOT_SUFFICIENT_MESSAGE,
+                )
+            ],
+        )
+
     compare_answers_prompt = INITIAL_REFINED_ANSWER_COMPARISON_PROMPT.format(
         question=question, initial_answer=initial_answer, refined_answer=refined_answer
     )
 
     msg = [HumanMessage(content=compare_answers_prompt)]
 
+    agent_error: AgentErrorLog | None = None
     # Get the rewritten queries in a defined format
     model = graph_config.tooling.fast_llm
-
+    resp: BaseMessage | None = None
+    refined_answer_improvement: bool | None = None
     # no need to stream this
-    resp = model.invoke(msg)
+    try:
+        resp = model.invoke(
+            msg, timeout_override=AGENT_TIMEOUT_OVERRIDE_LLM_COMPARE_ANSWERS
+        )
 
-    refined_answer_improvement = (
-        isinstance(resp.content, str) and "yes" in resp.content.lower()
-    )
+    except LLMTimeoutError:
+        agent_error = AgentErrorLog(
+            error_type=AgentLLMErrorType.TIMEOUT,
+            error_message=AGENT_LLM_TIMEOUT_MESSAGE,
+            error_result=_llm_node_error_strings.timeout,
+        )
+        logger.error("LLM Timeout Error - compare answers")
+        # continue as True in this support step
+    except LLMRateLimitError:
+        agent_error = AgentErrorLog(
+            error_type=AgentLLMErrorType.RATE_LIMIT,
+            error_message=AGENT_LLM_RATELIMIT_MESSAGE,
+            error_result=_llm_node_error_strings.rate_limit,
+        )
+        logger.error("LLM Rate Limit Error - compare answers")
+        # continue as True in this support step
+
+    if agent_error or resp is None:
+        refined_answer_improvement = True
+        if agent_error:
+            log_result = agent_error.error_result
+        else:
+            log_result = "An answer could not be generated."
+
+    else:
+        refined_answer_improvement = binary_string_test(
+            text=cast(str, resp.content),
+            positive_value=AGENT_POSITIVE_VALUE_STR,
+        )
+        log_result = f"Answer comparison: {refined_answer_improvement}"
 
     write_custom_event(
         "refined_answer_improvement",
@@ -65,7 +155,7 @@ def compare_answers(
                 graph_component="main",
                 node_name="compare answers",
                 node_start_time=node_start_time,
-                result=f"Answer comparison: {refined_answer_improvement}",
+                result=log_result,
             )
         ],
     )

backend/onyx/agents/agent_search/deep_search/main/nodes/create_refined_sub_questions.py

@@ -21,6 +21,18 @@ from onyx.agents.agent_search.models import GraphConfig
 from onyx.agents.agent_search.shared_graph_utils.agent_prompt_ops import (
     build_history_prompt,
 )
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AGENT_LLM_RATELIMIT_MESSAGE,
+)
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AGENT_LLM_TIMEOUT_MESSAGE,
+)
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AgentLLMErrorType,
+)
+from onyx.agents.agent_search.shared_graph_utils.models import AgentErrorLog
+from onyx.agents.agent_search.shared_graph_utils.models import BaseMessage_Content
+from onyx.agents.agent_search.shared_graph_utils.models import LLMNodeErrorStrings
 from onyx.agents.agent_search.shared_graph_utils.utils import dispatch_separated
 from onyx.agents.agent_search.shared_graph_utils.utils import (
     format_entity_term_extraction,
@@ -30,12 +42,31 @@ from onyx.agents.agent_search.shared_graph_utils.utils import (
 )
 from onyx.agents.agent_search.shared_graph_utils.utils import make_question_id
 from onyx.agents.agent_search.shared_graph_utils.utils import write_custom_event
+from onyx.chat.models import StreamingError
+from onyx.configs.agent_configs import (
+    AGENT_TIMEOUT_OVERRIDE_LLM_REFINED_SUBQUESTION_GENERATION,
+)
+from onyx.llm.chat_llm import LLMRateLimitError
+from onyx.llm.chat_llm import LLMTimeoutError
 from onyx.prompts.agent_search import (
-    REFINEMENT_QUESTION_DECOMPOSITION_PROMPT,
+    REFINEMENT_QUESTION_DECOMPOSITION_PROMPT_W_INITIAL_SUBQUESTION_ANSWERS,
 )
 from onyx.tools.models import ToolCallKickoff
+from onyx.utils.logger import setup_logger
+from onyx.utils.timing import log_function_time
+
+logger = setup_logger()
+
+_ANSWERED_SUBQUESTIONS_DIVIDER = "\n\n---\n\n"
+
+_llm_node_error_strings = LLMNodeErrorStrings(
+    timeout="The LLM timed out. The sub-questions could not be generated.",
+    rate_limit="The LLM encountered a rate limit. The sub-questions could not be generated.",
+    general_error="The LLM encountered an error. The sub-questions could not be generated.",
+)
 
 
+@log_function_time(print_only=True)
 def create_refined_sub_questions(
     state: MainState, config: RunnableConfig, writer: StreamWriter = lambda _: None
 ) -> RefinedQuestionDecompositionUpdate:
@@ -72,8 +103,10 @@ def create_refined_sub_questions(
 
     initial_question_answers = state.sub_question_results
 
-    addressed_question_list = [
-        x.question for x in initial_question_answers if x.verified_high_quality
+    addressed_subquestions_with_answers = [
+        f"Subquestion: {x.question}\nSubanswer:\n{x.answer}"
+        for x in initial_question_answers
+        if x.verified_high_quality and x.answer
     ]
 
     failed_question_list = [
@@ -82,12 +115,14 @@ def create_refined_sub_questions(
 
     msg = [
         HumanMessage(
-            content=REFINEMENT_QUESTION_DECOMPOSITION_PROMPT.format(
+            content=REFINEMENT_QUESTION_DECOMPOSITION_PROMPT_W_INITIAL_SUBQUESTION_ANSWERS.format(
                 question=question,
                 history=history,
                 entity_term_extraction_str=entity_term_extraction_str,
                 base_answer=base_answer,
-                answered_sub_questions="\n - ".join(addressed_question_list),
+                answered_subquestions_with_answers=_ANSWERED_SUBQUESTIONS_DIVIDER.join(
+                    addressed_subquestions_with_answers
+                ),
                 failed_sub_questions="\n - ".join(failed_question_list),
             ),
         )
@@ -96,29 +131,65 @@ def create_refined_sub_questions(
     # Grader
     model = graph_config.tooling.fast_llm
 
-    streamed_tokens = dispatch_separated(
-        model.stream(msg),
-        dispatch_subquestion(1, writer),
-        sep_callback=dispatch_subquestion_sep(1, writer),
-    )
-    response = merge_content(*streamed_tokens)
+    agent_error: AgentErrorLog | None = None
+    streamed_tokens: list[BaseMessage_Content] = []
+    try:
+        streamed_tokens = dispatch_separated(
+            model.stream(
+                msg,
+                timeout_override=AGENT_TIMEOUT_OVERRIDE_LLM_REFINED_SUBQUESTION_GENERATION,
+            ),
+            dispatch_subquestion(1, writer),
+            sep_callback=dispatch_subquestion_sep(1, writer),
+        )
+    except LLMTimeoutError:
+        agent_error = AgentErrorLog(
+            error_type=AgentLLMErrorType.TIMEOUT,
+            error_message=AGENT_LLM_TIMEOUT_MESSAGE,
+            error_result=_llm_node_error_strings.timeout,
+        )
+        logger.error("LLM Timeout Error - create refined sub questions")
 
-    if isinstance(response, str):
-        parsed_response = [q for q in response.split("\n") if q.strip() != ""]
-    else:
-        raise ValueError("LLM response is not a string")
+    except LLMRateLimitError:
+        agent_error = AgentErrorLog(
+            error_type=AgentLLMErrorType.RATE_LIMIT,
+            error_message=AGENT_LLM_RATELIMIT_MESSAGE,
+            error_result=_llm_node_error_strings.rate_limit,
+        )
+        logger.error("LLM Rate Limit Error - create refined sub questions")
 
-    refined_sub_question_dict = {}
-    for sub_question_num, sub_question in enumerate(parsed_response):
-        refined_sub_question = RefinementSubQuestion(
-            sub_question=sub_question,
-            sub_question_id=make_question_id(1, sub_question_num + 1),
-            verified=False,
-            answered=False,
-            answer="",
+    if agent_error:
+        refined_sub_question_dict: dict[int, RefinementSubQuestion] = {}
+        log_result = agent_error.error_result
+        write_custom_event(
+            "refined_sub_question_creation_error",
+            StreamingError(
+                error="Your LLM was not able to create refined sub questions in time and timed out. Please try again.",
+            ),
+            writer,
         )
 
-        refined_sub_question_dict[sub_question_num + 1] = refined_sub_question
+    else:
+        response = merge_content(*streamed_tokens)
+
+        if isinstance(response, str):
+            parsed_response = [q for q in response.split("\n") if q.strip() != ""]
+        else:
+            raise ValueError("LLM response is not a string")
+
+        refined_sub_question_dict = {}
+        for sub_question_num, sub_question in enumerate(parsed_response):
+            refined_sub_question = RefinementSubQuestion(
+                sub_question=sub_question,
+                sub_question_id=make_question_id(1, sub_question_num + 1),
+                verified=False,
+                answered=False,
+                answer="",
+            )
+
+            refined_sub_question_dict[sub_question_num + 1] = refined_sub_question
+
+        log_result = f"Created {len(refined_sub_question_dict)} refined sub questions"
 
     return RefinedQuestionDecompositionUpdate(
         refined_sub_questions=refined_sub_question_dict,
@@ -128,7 +199,7 @@ def create_refined_sub_questions(
                 graph_component="main",
                 node_name="create refined sub questions",
                 node_start_time=node_start_time,
-                result=f"Created {len(refined_sub_question_dict)} refined sub questions",
+                result=log_result,
             )
         ],
     )

backend/onyx/agents/agent_search/deep_search/main/nodes/decide_refinement_need.py

@@ -11,8 +11,10 @@ from onyx.agents.agent_search.models import GraphConfig
 from onyx.agents.agent_search.shared_graph_utils.utils import (
     get_langgraph_node_log_string,
 )
+from onyx.utils.timing import log_function_time
 
 
+@log_function_time(print_only=True)
 def decide_refinement_need(
     state: MainState, config: RunnableConfig
 ) -> RequireRefinemenEvalUpdate:
@@ -26,6 +28,19 @@ def decide_refinement_need(
 
     decision = True  # TODO: just for current testing purposes
 
+    if state.answer_error:
+        return RequireRefinemenEvalUpdate(
+            require_refined_answer_eval=False,
+            log_messages=[
+                get_langgraph_node_log_string(
+                    graph_component="main",
+                    node_name="decide refinement need",
+                    node_start_time=node_start_time,
+                    result="Timeout Error",
+                )
+            ],
+        )
+
     log_messages = [
         get_langgraph_node_log_string(
             graph_component="main",

backend/onyx/agents/agent_search/deep_search/main/nodes/extract_entities_terms.py

@@ -21,11 +21,16 @@ from onyx.agents.agent_search.shared_graph_utils.utils import format_docs
 from onyx.agents.agent_search.shared_graph_utils.utils import (
     get_langgraph_node_log_string,
 )
+from onyx.configs.agent_configs import (
+    AGENT_TIMEOUT_OVERRIDE_LLM_ENTITY_TERM_EXTRACTION,
+)
 from onyx.configs.constants import NUM_EXPLORATORY_DOCS
 from onyx.prompts.agent_search import ENTITY_TERM_EXTRACTION_PROMPT
 from onyx.prompts.agent_search import ENTITY_TERM_EXTRACTION_PROMPT_JSON_EXAMPLE
+from onyx.utils.timing import log_function_time
 
 
+@log_function_time(print_only=True)
 def extract_entities_terms(
     state: MainState, config: RunnableConfig
 ) -> EntityTermExtractionUpdate:
@@ -81,6 +86,7 @@ def extract_entities_terms(
     # Grader
     llm_response = fast_llm.invoke(
         prompt=msg,
+        timeout_override=AGENT_TIMEOUT_OVERRIDE_LLM_ENTITY_TERM_EXTRACTION,
     )
 
     cleaned_response = (

backend/onyx/agents/agent_search/deep_search/main/nodes/generate_validate_refined_answer.py

@@ -11,27 +11,49 @@ from onyx.agents.agent_search.deep_search.main.models import (
     AgentRefinedMetrics,
 )
 from onyx.agents.agent_search.deep_search.main.operations import get_query_info
-from onyx.agents.agent_search.deep_search.main.operations import logger
 from onyx.agents.agent_search.deep_search.main.states import MainState
 from onyx.agents.agent_search.deep_search.main.states import (
     RefinedAnswerUpdate,
 )
 from onyx.agents.agent_search.models import GraphConfig
+from onyx.agents.agent_search.shared_graph_utils.agent_prompt_ops import (
+    binary_string_test_after_answer_separator,
+)
 from onyx.agents.agent_search.shared_graph_utils.agent_prompt_ops import (
     get_prompt_enrichment_components,
 )
 from onyx.agents.agent_search.shared_graph_utils.agent_prompt_ops import (
     trim_prompt_piece,
 )
-from onyx.agents.agent_search.shared_graph_utils.models import InferenceSection
+from onyx.agents.agent_search.shared_graph_utils.calculations import (
+    get_answer_generation_documents,
+)
+from onyx.agents.agent_search.shared_graph_utils.constants import AGENT_ANSWER_SEPARATOR
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AGENT_LLM_RATELIMIT_MESSAGE,
+)
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AGENT_LLM_TIMEOUT_MESSAGE,
+)
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AGENT_POSITIVE_VALUE_STR,
+)
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AgentLLMErrorType,
+)
+from onyx.agents.agent_search.shared_graph_utils.models import AgentErrorLog
+from onyx.agents.agent_search.shared_graph_utils.models import LLMNodeErrorStrings
 from onyx.agents.agent_search.shared_graph_utils.models import RefinedAgentStats
 from onyx.agents.agent_search.shared_graph_utils.operators import (
-    dedup_inference_sections,
+    dedup_inference_section_list,
 )
 from onyx.agents.agent_search.shared_graph_utils.utils import (
     dispatch_main_answer_stop_info,
 )
 from onyx.agents.agent_search.shared_graph_utils.utils import format_docs
+from onyx.agents.agent_search.shared_graph_utils.utils import (
+    get_deduplicated_structured_subquestion_documents,
+)
 from onyx.agents.agent_search.shared_graph_utils.utils import (
     get_langgraph_node_log_string,
 )
@@ -43,26 +65,50 @@ from onyx.agents.agent_search.shared_graph_utils.utils import (
 from onyx.agents.agent_search.shared_graph_utils.utils import write_custom_event
 from onyx.chat.models import AgentAnswerPiece
 from onyx.chat.models import ExtendedToolResponse
+from onyx.chat.models import StreamingError
 from onyx.configs.agent_configs import AGENT_MAX_ANSWER_CONTEXT_DOCS
+from onyx.configs.agent_configs import AGENT_MAX_STREAMED_DOCS_FOR_REFINED_ANSWER
 from onyx.configs.agent_configs import AGENT_MIN_ORIG_QUESTION_DOCS
+from onyx.configs.agent_configs import (
+    AGENT_TIMEOUT_OVERRIDE_LLM_REFINED_ANSWER_GENERATION,
+)
+from onyx.configs.agent_configs import (
+    AGENT_TIMEOUT_OVERRIDE_LLM_REFINED_ANSWER_VALIDATION,
+)
+from onyx.llm.chat_llm import LLMRateLimitError
+from onyx.llm.chat_llm import LLMTimeoutError
 from onyx.prompts.agent_search import (
     REFINED_ANSWER_PROMPT_W_SUB_QUESTIONS,
 )
 from onyx.prompts.agent_search import (
     REFINED_ANSWER_PROMPT_WO_SUB_QUESTIONS,
 )
+from onyx.prompts.agent_search import (
+    REFINED_ANSWER_VALIDATION_PROMPT,
+)
 from onyx.prompts.agent_search import (
     SUB_QUESTION_ANSWER_TEMPLATE_REFINED,
 )
 from onyx.prompts.agent_search import UNKNOWN_ANSWER
 from onyx.tools.tool_implementations.search.search_tool import yield_search_responses
+from onyx.utils.logger import setup_logger
+from onyx.utils.timing import log_function_time
+
+logger = setup_logger()
+
+_llm_node_error_strings = LLMNodeErrorStrings(
+    timeout="The LLM timed out. The refined answer could not be generated.",
+    rate_limit="The LLM encountered a rate limit. The refined answer could not be generated.",
+    general_error="The LLM encountered an error. The refined answer could not be generated.",
+)
 
 
-def generate_refined_answer(
+@log_function_time(print_only=True)
+def generate_validate_refined_answer(
     state: MainState, config: RunnableConfig, writer: StreamWriter = lambda _: None
 ) -> RefinedAnswerUpdate:
     """
-    LangGraph node to generate the refined answer.
+    LangGraph node to generate the refined answer and validate it.
     """
 
     node_start_time = datetime.now()
@@ -76,19 +122,24 @@ def generate_refined_answer(
     )
 
     verified_reranked_documents = state.verified_reranked_documents
-    sub_questions_cited_documents = state.cited_documents
+
+    # get all documents cited in sub-questions
+    structured_subquestion_docs = get_deduplicated_structured_subquestion_documents(
+        state.sub_question_results
+    )
+
     original_question_verified_documents = (
         state.orig_question_verified_reranked_documents
     )
     original_question_retrieved_documents = state.orig_question_retrieved_documents
 
-    consolidated_context_docs: list[InferenceSection] = sub_questions_cited_documents
+    consolidated_context_docs = structured_subquestion_docs.cited_documents
 
     counter = 0
     for original_doc_number, original_doc in enumerate(
         original_question_verified_documents
     ):
-        if original_doc_number not in sub_questions_cited_documents:
+        if original_doc_number not in structured_subquestion_docs.cited_documents:
             if (
                 counter <= AGENT_MIN_ORIG_QUESTION_DOCS
                 or len(consolidated_context_docs)
@@ -99,14 +150,16 @@ def generate_refined_answer(
                 counter += 1
 
     # sort docs by their scores - though the scores refer to different questions
-    relevant_docs = dedup_inference_sections(
-        consolidated_context_docs, consolidated_context_docs
-    )
+    relevant_docs = dedup_inference_section_list(consolidated_context_docs)
 
-    streaming_docs = (
-        relevant_docs
-        if len(relevant_docs) > 0
-        else original_question_retrieved_documents[:15]
+    # Create the list of documents to stream out. Start with the
+    # ones that wil be in the context (or, if len == 0, use docs
+    # that were retrieved for the original question)
+    answer_generation_documents = get_answer_generation_documents(
+        relevant_docs=relevant_docs,
+        context_documents=structured_subquestion_docs.context_documents,
+        original_question_docs=original_question_retrieved_documents,
+        max_docs=AGENT_MAX_STREAMED_DOCS_FOR_REFINED_ANSWER,
     )
 
     query_info = get_query_info(state.orig_question_sub_query_retrieval_results)
@@ -114,11 +167,13 @@ def generate_refined_answer(
         graph_config.tooling.search_tool
     ), "search_tool must be provided for agentic search"
     # stream refined answer docs, or original question docs if no relevant docs are found
-    relevance_list = relevance_from_docs(relevant_docs)
+    relevance_list = relevance_from_docs(
+        answer_generation_documents.streaming_documents
+    )
     for tool_response in yield_search_responses(
         query=question,
-        reranked_sections=streaming_docs,
-        final_context_sections=streaming_docs,
+        reranked_sections=answer_generation_documents.streaming_documents,
+        final_context_sections=answer_generation_documents.context_documents,
         search_query_info=query_info,
         get_section_relevance=lambda: relevance_list,
         search_tool=graph_config.tooling.search_tool,
@@ -199,7 +254,7 @@ def generate_refined_answer(
     )
 
     model = graph_config.tooling.fast_llm
-    relevant_docs_str = format_docs(relevant_docs)
+    relevant_docs_str = format_docs(answer_generation_documents.context_documents)
     relevant_docs_str = trim_prompt_piece(
         model.config,
         relevant_docs_str,
@@ -231,28 +286,80 @@ def generate_refined_answer(
 
     streamed_tokens: list[str | list[str | dict[str, Any]]] = [""]
     dispatch_timings: list[float] = []
-    for message in model.stream(msg):
-        # TODO: in principle, the answer here COULD contain images, but we don't support that yet
-        content = message.content
-        if not isinstance(content, str):
-            raise ValueError(
-                f"Expected content to be a string, but got {type(content)}"
-            )
+    agent_error: AgentErrorLog | None = None
 
-        start_stream_token = datetime.now()
+    try:
+        for message in model.stream(
+            msg, timeout_override=AGENT_TIMEOUT_OVERRIDE_LLM_REFINED_ANSWER_GENERATION
+        ):
+            # TODO: in principle, the answer here COULD contain images, but we don't support that yet
+            content = message.content
+            if not isinstance(content, str):
+                raise ValueError(
+                    f"Expected content to be a string, but got {type(content)}"
+                )
+
+            start_stream_token = datetime.now()
+            write_custom_event(
+                "refined_agent_answer",
+                AgentAnswerPiece(
+                    answer_piece=content,
+                    level=1,
+                    level_question_num=0,
+                    answer_type="agent_level_answer",
+                ),
+                writer,
+            )
+            end_stream_token = datetime.now()
+            dispatch_timings.append(
+                (end_stream_token - start_stream_token).microseconds
+            )
+            streamed_tokens.append(content)
+
+    except LLMTimeoutError:
+        agent_error = AgentErrorLog(
+            error_type=AgentLLMErrorType.TIMEOUT,
+            error_message=AGENT_LLM_TIMEOUT_MESSAGE,
+            error_result=_llm_node_error_strings.timeout,
+        )
+        logger.error("LLM Timeout Error - generate refined answer")
+
+    except LLMRateLimitError:
+        agent_error = AgentErrorLog(
+            error_type=AgentLLMErrorType.RATE_LIMIT,
+            error_message=AGENT_LLM_RATELIMIT_MESSAGE,
+            error_result=_llm_node_error_strings.rate_limit,
+        )
+        logger.error("LLM Rate Limit Error - generate refined answer")
+
+    if agent_error:
         write_custom_event(
-            "refined_agent_answer",
-            AgentAnswerPiece(
-                answer_piece=content,
-                level=1,
-                level_question_num=0,
-                answer_type="agent_level_answer",
+            "initial_agent_answer",
+            StreamingError(
+                error=AGENT_LLM_TIMEOUT_MESSAGE,
             ),
             writer,
         )
-        end_stream_token = datetime.now()
-        dispatch_timings.append((end_stream_token - start_stream_token).microseconds)
-        streamed_tokens.append(content)
+
+        return RefinedAnswerUpdate(
+            refined_answer=None,
+            refined_answer_quality=False,  # TODO: replace this with the actual check value
+            refined_agent_stats=None,
+            agent_refined_end_time=None,
+            agent_refined_metrics=AgentRefinedMetrics(
+                refined_doc_boost_factor=0.0,
+                refined_question_boost_factor=0.0,
+                duration_s=None,
+            ),
+            log_messages=[
+                get_langgraph_node_log_string(
+                    graph_component="main",
+                    node_name="generate refined answer",
+                    node_start_time=node_start_time,
+                    result=agent_error.error_result or "An LLM error occurred",
+                )
+            ],
+        )
 
     logger.debug(
         f"Average dispatch time for refined answer: {sum(dispatch_timings) / len(dispatch_timings)}"
@@ -261,54 +368,43 @@ def generate_refined_answer(
     response = merge_content(*streamed_tokens)
     answer = cast(str, response)
 
+    # run a validation step for the refined answer only
+
+    msg = [
+        HumanMessage(
+            content=REFINED_ANSWER_VALIDATION_PROMPT.format(
+                question=question,
+                history=prompt_enrichment_components.history,
+                answered_sub_questions=sub_question_answer_str,
+                relevant_docs=relevant_docs_str,
+                proposed_answer=answer,
+                persona_specification=persona_contextualized_prompt,
+            )
+        )
+    ]
+
+    try:
+        validation_response = model.invoke(
+            msg, timeout_override=AGENT_TIMEOUT_OVERRIDE_LLM_REFINED_ANSWER_VALIDATION
+        )
+        refined_answer_quality = binary_string_test_after_answer_separator(
+            text=cast(str, validation_response.content),
+            positive_value=AGENT_POSITIVE_VALUE_STR,
+            separator=AGENT_ANSWER_SEPARATOR,
+        )
+    except LLMTimeoutError:
+        refined_answer_quality = True
+        logger.error("LLM Timeout Error - validate refined answer")
+
+    except LLMRateLimitError:
+        refined_answer_quality = True
+        logger.error("LLM Rate Limit Error - validate refined answer")
+
     refined_agent_stats = RefinedAgentStats(
         revision_doc_efficiency=refined_doc_effectiveness,
         revision_question_efficiency=revision_question_efficiency,
     )
 
-    logger.debug(f"\n\n---INITIAL ANSWER ---\n\n Answer:\n Agent: {initial_answer}")
-    logger.debug("-" * 10)
-    logger.debug(f"\n\n---REVISED AGENT ANSWER ---\n\n Answer:\n Agent: {answer}")
-
-    logger.debug("-" * 100)
-
-    if state.initial_agent_stats:
-        initial_doc_boost_factor = state.initial_agent_stats.agent_effectiveness.get(
-            "utilized_chunk_ratio", "--"
-        )
-        initial_support_boost_factor = (
-            state.initial_agent_stats.agent_effectiveness.get("support_ratio", "--")
-        )
-        num_initial_verified_docs = state.initial_agent_stats.original_question.get(
-            "num_verified_documents", "--"
-        )
-        initial_verified_docs_avg_score = (
-            state.initial_agent_stats.original_question.get("verified_avg_score", "--")
-        )
-        initial_sub_questions_verified_docs = (
-            state.initial_agent_stats.sub_questions.get("num_verified_documents", "--")
-        )
-
-        logger.debug("INITIAL AGENT STATS")
-        logger.debug(f"Document Boost Factor: {initial_doc_boost_factor}")
-        logger.debug(f"Support Boost Factor: {initial_support_boost_factor}")
-        logger.debug(f"Originally Verified Docs: {num_initial_verified_docs}")
-        logger.debug(
-            f"Originally Verified Docs Avg Score: {initial_verified_docs_avg_score}"
-        )
-        logger.debug(
-            f"Sub-Questions Verified Docs: {initial_sub_questions_verified_docs}"
-        )
-    if refined_agent_stats:
-        logger.debug("-" * 10)
-        logger.debug("REFINED AGENT STATS")
-        logger.debug(
-            f"Revision Doc Factor: {refined_agent_stats.revision_doc_efficiency}"
-        )
-        logger.debug(
-            f"Revision Question Factor: {refined_agent_stats.revision_question_efficiency}"
-        )
-
     agent_refined_end_time = datetime.now()
     if state.agent_refined_start_time:
         agent_refined_duration = (
@@ -325,7 +421,7 @@ def generate_refined_answer(
 
     return RefinedAnswerUpdate(
         refined_answer=answer,
-        refined_answer_quality=True,  # TODO: replace this with the actual check value
+        refined_answer_quality=refined_answer_quality,
         refined_agent_stats=refined_agent_stats,
         agent_refined_end_time=agent_refined_end_time,
         agent_refined_metrics=agent_refined_metrics,

backend/onyx/agents/agent_search/deep_search/main/states.py

@@ -17,6 +17,7 @@ from onyx.agents.agent_search.orchestration.states import ToolCallUpdate
 from onyx.agents.agent_search.orchestration.states import ToolChoiceInput
 from onyx.agents.agent_search.orchestration.states import ToolChoiceUpdate
 from onyx.agents.agent_search.shared_graph_utils.models import AgentChunkRetrievalStats
+from onyx.agents.agent_search.shared_graph_utils.models import AgentErrorLog
 from onyx.agents.agent_search.shared_graph_utils.models import (
     EntityRelationshipTermExtraction,
 )
@@ -76,6 +77,7 @@ class InitialAnswerUpdate(LoggerUpdate):
     """
 
     initial_answer: str | None = None
+    answer_error: AgentErrorLog | None = None
     initial_agent_stats: InitialAgentResultStats | None = None
     generated_sub_questions: list[str] = []
     agent_base_end_time: datetime | None = None
@@ -88,6 +90,7 @@ class RefinedAnswerUpdate(RefinedAgentEndStats, LoggerUpdate):
     """
 
     refined_answer: str | None = None
+    answer_error: AgentErrorLog | None = None
     refined_agent_stats: RefinedAgentStats | None = None
     refined_answer_quality: bool = False
 

backend/onyx/agents/agent_search/deep_search/shared/expanded_retrieval/nodes/expand_queries.py

@@ -16,16 +16,44 @@ from onyx.agents.agent_search.deep_search.shared.expanded_retrieval.states impor
     QueryExpansionUpdate,
 )
 from onyx.agents.agent_search.models import GraphConfig
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AGENT_LLM_RATELIMIT_MESSAGE,
+)
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AGENT_LLM_TIMEOUT_MESSAGE,
+)
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AgentLLMErrorType,
+)
+from onyx.agents.agent_search.shared_graph_utils.models import AgentErrorLog
+from onyx.agents.agent_search.shared_graph_utils.models import BaseMessage_Content
+from onyx.agents.agent_search.shared_graph_utils.models import LLMNodeErrorStrings
 from onyx.agents.agent_search.shared_graph_utils.utils import dispatch_separated
 from onyx.agents.agent_search.shared_graph_utils.utils import (
     get_langgraph_node_log_string,
 )
 from onyx.agents.agent_search.shared_graph_utils.utils import parse_question_id
+from onyx.configs.agent_configs import (
+    AGENT_TIMEOUT_OVERRIDE_LLM_QUERY_REWRITING_GENERATION,
+)
+from onyx.llm.chat_llm import LLMRateLimitError
+from onyx.llm.chat_llm import LLMTimeoutError
 from onyx.prompts.agent_search import (
     QUERY_REWRITING_PROMPT,
 )
+from onyx.utils.logger import setup_logger
+from onyx.utils.timing import log_function_time
+
+logger = setup_logger()
+
+_llm_node_error_strings = LLMNodeErrorStrings(
+    timeout="Query rewriting failed due to LLM timeout - the original question will be used.",
+    rate_limit="Query rewriting failed due to LLM rate limit - the original question will be used.",
+    general_error="Query rewriting failed due to LLM error - the original question will be used.",
+)
 
 
+@log_function_time(print_only=True)
 def expand_queries(
     state: ExpandedRetrievalInput,
     config: RunnableConfig,
@@ -54,13 +82,43 @@ def expand_queries(
         )
     ]
 
-    llm_response_list = dispatch_separated(
-        llm.stream(prompt=msg), dispatch_subquery(level, question_num, writer)
-    )
+    agent_error: AgentErrorLog | None = None
+    llm_response_list: list[BaseMessage_Content] = []
+    llm_response = ""
+    rewritten_queries = []
 
-    llm_response = merge_message_runs(llm_response_list, chunk_separator="")[0].content
+    try:
+        llm_response_list = dispatch_separated(
+            llm.stream(
+                prompt=msg,
+                timeout_override=AGENT_TIMEOUT_OVERRIDE_LLM_QUERY_REWRITING_GENERATION,
+            ),
+            dispatch_subquery(level, question_num, writer),
+        )
+        llm_response = merge_message_runs(llm_response_list, chunk_separator="")[
+            0
+        ].content
+        rewritten_queries = llm_response.split("\n")
+        log_result = f"Number of expanded queries: {len(rewritten_queries)}"
 
-    rewritten_queries = llm_response.split("\n")
+    except LLMTimeoutError:
+        agent_error = AgentErrorLog(
+            error_type=AgentLLMErrorType.TIMEOUT,
+            error_message=AGENT_LLM_TIMEOUT_MESSAGE,
+            error_result=_llm_node_error_strings.timeout,
+        )
+        logger.error("LLM Timeout Error - expand queries")
+        log_result = agent_error.error_result
+
+    except LLMRateLimitError:
+        agent_error = AgentErrorLog(
+            error_type=AgentLLMErrorType.RATE_LIMIT,
+            error_message=AGENT_LLM_RATELIMIT_MESSAGE,
+            error_result=_llm_node_error_strings.rate_limit,
+        )
+        logger.error("LLM Rate Limit Error - expand queries")
+        log_result = agent_error.error_result
+    # use subquestion as query if query generation fails
 
     return QueryExpansionUpdate(
         expanded_queries=rewritten_queries,
@@ -69,7 +127,7 @@ def expand_queries(
                 graph_component="shared - expanded retrieval",
                 node_name="expand queries",
                 node_start_time=node_start_time,
-                result=f"Number of expanded queries: {len(rewritten_queries)}",
+                result=log_result,
             )
         ],
     )

backend/onyx/agents/agent_search/deep_search/shared/expanded_retrieval/nodes/rerank_documents.py

@@ -21,12 +21,15 @@ from onyx.agents.agent_search.shared_graph_utils.utils import (
 from onyx.configs.agent_configs import AGENT_RERANKING_MAX_QUERY_RETRIEVAL_RESULTS
 from onyx.configs.agent_configs import AGENT_RERANKING_STATS
 from onyx.context.search.models import InferenceSection
-from onyx.context.search.models import SearchRequest
-from onyx.context.search.pipeline import retrieval_preprocessing
+from onyx.context.search.models import RerankingDetails
 from onyx.context.search.postprocessing.postprocessing import rerank_sections
+from onyx.context.search.postprocessing.postprocessing import should_rerank
 from onyx.db.engine import get_session_context_manager
+from onyx.db.search_settings import get_current_search_settings
+from onyx.utils.timing import log_function_time
 
 
+@log_function_time(print_only=True)
 def rerank_documents(
     state: ExpandedRetrievalState, config: RunnableConfig
 ) -> DocRerankingUpdate:
@@ -39,6 +42,8 @@ def rerank_documents(
 
     # Rerank post retrieval and verification. First, create a search query
     # then create the list of reranked sections
+    # If no question defined/question is None in the state, use the original
+    # question from the search request as query
 
     graph_config = cast(GraphConfig, config["metadata"]["config"])
     question = (
@@ -47,39 +52,28 @@ def rerank_documents(
     assert (
         graph_config.tooling.search_tool
     ), "search_tool must be provided for agentic search"
-    with get_session_context_manager() as db_session:
-        # we ignore some of the user specified fields since this search is
-        # internal to agentic search, but we still want to pass through
-        # persona (for stuff like document sets) and rerank settings
-        # (to not make an unnecessary db call).
-        search_request = SearchRequest(
-            query=question,
-            persona=graph_config.inputs.search_request.persona,
-            rerank_settings=graph_config.inputs.search_request.rerank_settings,
-        )
-        _search_query = retrieval_preprocessing(
-            search_request=search_request,
-            user=graph_config.tooling.search_tool.user,  # bit of a hack
-            llm=graph_config.tooling.fast_llm,
-            db_session=db_session,
-        )
 
-    # skip section filtering
+    # Note that these are passed in values from the API and are overrides which are typically None
+    rerank_settings = graph_config.inputs.search_request.rerank_settings
 
-    if (
-        _search_query.rerank_settings
-        and _search_query.rerank_settings.rerank_model_name
-        and _search_query.rerank_settings.num_rerank > 0
-        and len(verified_documents) > 0
-    ):
+    if rerank_settings is None:
+        with get_session_context_manager() as db_session:
+            search_settings = get_current_search_settings(db_session)
+            if not search_settings.disable_rerank_for_streaming:
+                rerank_settings = RerankingDetails.from_db_model(search_settings)
+
+    if should_rerank(rerank_settings) and len(verified_documents) > 0:
         if len(verified_documents) > 1:
             reranked_documents = rerank_sections(
-                _search_query,
-                verified_documents,
+                query_str=question,
+                # if runnable, then rerank_settings is not None
+                rerank_settings=cast(RerankingDetails, rerank_settings),
+                sections_to_rerank=verified_documents,
             )
         else:
-            num = "No" if len(verified_documents) == 0 else "One"
-            logger.warning(f"{num} verified document(s) found, skipping reranking")
+            logger.warning(
+                f"{len(verified_documents)} verified document(s) found, skipping reranking"
+            )
             reranked_documents = verified_documents
     else:
         logger.warning("No reranking settings found, using unranked documents")

backend/onyx/agents/agent_search/deep_search/shared/expanded_retrieval/nodes/retrieve_documents.py

@@ -23,12 +23,15 @@ from onyx.configs.agent_configs import AGENT_RETRIEVAL_STATS
 from onyx.context.search.models import InferenceSection
 from onyx.db.engine import get_session_context_manager
 from onyx.tools.models import SearchQueryInfo
+from onyx.tools.models import SearchToolOverrideKwargs
 from onyx.tools.tool_implementations.search.search_tool import (
     SEARCH_RESPONSE_SUMMARY_ID,
 )
 from onyx.tools.tool_implementations.search.search_tool import SearchResponseSummary
+from onyx.utils.timing import log_function_time
 
 
+@log_function_time(print_only=True)
 def retrieve_documents(
     state: RetrievalInput, config: RunnableConfig
 ) -> DocRetrievalUpdate:
@@ -67,9 +70,12 @@ def retrieve_documents(
     with get_session_context_manager() as db_session:
         for tool_response in search_tool.run(
             query=query_to_retrieve,
-            force_no_rerank=True,
-            alternate_db_session=db_session,
-            retrieved_sections_callback=callback_container.append,
+            override_kwargs=SearchToolOverrideKwargs(
+                force_no_rerank=True,
+                alternate_db_session=db_session,
+                retrieved_sections_callback=callback_container.append,
+                skip_query_analysis=not state.base_search,
+            ),
         ):
             # get retrieved docs to send to the rest of the graph
             if tool_response.id == SEARCH_RESPONSE_SUMMARY_ID:

backend/onyx/agents/agent_search/deep_search/shared/expanded_retrieval/nodes/verify_documents.py

@@ -1,5 +1,7 @@
+from datetime import datetime
 from typing import cast
 
+from langchain_core.messages import BaseMessage
 from langchain_core.messages import HumanMessage
 from langchain_core.runnables.config import RunnableConfig
 
@@ -10,14 +12,38 @@ from onyx.agents.agent_search.deep_search.shared.expanded_retrieval.states impor
     DocVerificationUpdate,
 )
 from onyx.agents.agent_search.models import GraphConfig
+from onyx.agents.agent_search.shared_graph_utils.agent_prompt_ops import (
+    binary_string_test,
+)
 from onyx.agents.agent_search.shared_graph_utils.agent_prompt_ops import (
     trim_prompt_piece,
 )
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AGENT_POSITIVE_VALUE_STR,
+)
+from onyx.agents.agent_search.shared_graph_utils.models import LLMNodeErrorStrings
+from onyx.agents.agent_search.shared_graph_utils.utils import (
+    get_langgraph_node_log_string,
+)
+from onyx.configs.agent_configs import AGENT_TIMEOUT_OVERRIDE_LLM_DOCUMENT_VERIFICATION
+from onyx.llm.chat_llm import LLMRateLimitError
+from onyx.llm.chat_llm import LLMTimeoutError
 from onyx.prompts.agent_search import (
     DOCUMENT_VERIFICATION_PROMPT,
 )
+from onyx.utils.logger import setup_logger
+from onyx.utils.timing import log_function_time
+
+logger = setup_logger()
+
+_llm_node_error_strings = LLMNodeErrorStrings(
+    timeout="The LLM timed out. The document could not be verified. The document will be treated as 'relevant'",
+    rate_limit="The LLM encountered a rate limit. The document could not be verified. The document will be treated as 'relevant'",
+    general_error="The LLM encountered an error. The document could not be verified. The document will be treated as 'relevant'",
+)
 
 
+@log_function_time(print_only=True)
 def verify_documents(
     state: DocVerificationInput, config: RunnableConfig
 ) -> DocVerificationUpdate:
@@ -26,12 +52,14 @@ def verify_documents(
 
     Args:
         state (DocVerificationInput): The current state
-        config (RunnableConfig): Configuration containing ProSearchConfig
+        config (RunnableConfig): Configuration containing AgentSearchConfig
 
     Updates:
         verified_documents: list[InferenceSection]
     """
 
+    node_start_time = datetime.now()
+
     question = state.question
     retrieved_document_to_verify = state.retrieved_document_to_verify
     document_content = retrieved_document_to_verify.combined_content
@@ -51,12 +79,40 @@ def verify_documents(
         )
     ]
 
-    response = fast_llm.invoke(msg)
+    response: BaseMessage | None = None
 
-    verified_documents = []
-    if isinstance(response.content, str) and "yes" in response.content.lower():
-        verified_documents.append(retrieved_document_to_verify)
+    verified_documents = [
+        retrieved_document_to_verify
+    ]  # default is to treat document as relevant
+
+    try:
+        response = fast_llm.invoke(
+            msg, timeout_override=AGENT_TIMEOUT_OVERRIDE_LLM_DOCUMENT_VERIFICATION
+        )
+
+        assert isinstance(response.content, str)
+        if not binary_string_test(
+            text=response.content, positive_value=AGENT_POSITIVE_VALUE_STR
+        ):
+            verified_documents = []
+
+    except LLMTimeoutError:
+        # In this case, we decide to continue and don't raise an error, as
+        # little harm in letting some docs through that are less relevant.
+        logger.error("LLM Timeout Error - verify documents")
+
+    except LLMRateLimitError:
+        # In this case, we decide to continue and don't raise an error, as
+        # little harm in letting some docs through that are less relevant.
+        logger.error("LLM Rate Limit Error - verify documents")
 
     return DocVerificationUpdate(
         verified_documents=verified_documents,
+        log_messages=[
+            get_langgraph_node_log_string(
+                graph_component="shared - expanded retrieval",
+                node_name="verify documents",
+                node_start_time=node_start_time,
+            )
+        ],
     )

backend/onyx/agents/agent_search/deep_search/shared/expanded_retrieval/states.py

@@ -21,9 +21,13 @@ from onyx.context.search.models import InferenceSection
 
 
 class ExpandedRetrievalInput(SubgraphCoreState):
-    question: str = ""
-    base_search: bool = False
+    # exception from 'no default value'for LangGraph input states
+    # Here, sub_question_id default None implies usage for the
+    # original question. This is sometimes needed for nested sub-graphs
+
     sub_question_id: str | None = None
+    question: str
+    base_search: bool
 
 
 ## Update/Return States
@@ -34,7 +38,7 @@ class QueryExpansionUpdate(LoggerUpdate, BaseModel):
     log_messages: list[str] = []
 
 
-class DocVerificationUpdate(BaseModel):
+class DocVerificationUpdate(LoggerUpdate, BaseModel):
     verified_documents: Annotated[list[InferenceSection], dedup_inference_sections] = []
 
 
@@ -88,4 +92,4 @@ class DocVerificationInput(ExpandedRetrievalInput):
 
 
 class RetrievalInput(ExpandedRetrievalInput):
-    query_to_retrieve: str = ""
+    query_to_retrieve: str

backend/onyx/agents/agent_search/run_graph.py

@@ -12,7 +12,7 @@ from onyx.agents.agent_search.deep_search.main.graph_builder import (
     main_graph_builder as main_graph_builder_a,
 )
 from onyx.agents.agent_search.deep_search.main.states import (
-    MainInput as MainInput_a,
+    MainInput as MainInput,
 )
 from onyx.agents.agent_search.models import GraphConfig
 from onyx.agents.agent_search.shared_graph_utils.utils import get_test_config
@@ -21,6 +21,7 @@ from onyx.chat.models import AnswerPacket
 from onyx.chat.models import AnswerStream
 from onyx.chat.models import ExtendedToolResponse
 from onyx.chat.models import RefinedAnswerImprovement
+from onyx.chat.models import StreamingError
 from onyx.chat.models import StreamStopInfo
 from onyx.chat.models import SubQueryPiece
 from onyx.chat.models import SubQuestionPiece
@@ -33,6 +34,7 @@ from onyx.llm.factory import get_default_llms
 from onyx.tools.tool_runner import ToolCallKickoff
 from onyx.utils.logger import setup_logger
 
+
 logger = setup_logger()
 
 _COMPILED_GRAPH: CompiledStateGraph | None = None
@@ -72,13 +74,15 @@ def _parse_agent_event(
             return cast(AnswerPacket, event["data"])
         elif event["name"] == "refined_answer_improvement":
             return cast(RefinedAnswerImprovement, event["data"])
+        elif event["name"] == "refined_sub_question_creation_error":
+            return cast(StreamingError, event["data"])
     return None
 
 
 def manage_sync_streaming(
     compiled_graph: CompiledStateGraph,
     config: GraphConfig,
-    graph_input: BasicInput | MainInput_a,
+    graph_input: BasicInput | MainInput,
 ) -> Iterable[StreamEvent]:
     message_id = config.persistence.message_id if config.persistence else None
     for event in compiled_graph.stream(
@@ -92,7 +96,7 @@ def manage_sync_streaming(
 def run_graph(
     compiled_graph: CompiledStateGraph,
     config: GraphConfig,
-    input: BasicInput | MainInput_a,
+    input: BasicInput | MainInput,
 ) -> AnswerStream:
     config.behavior.perform_initial_search_decomposition = (
         INITIAL_SEARCH_DECOMPOSITION_ENABLED
@@ -123,9 +127,7 @@ def run_main_graph(
 ) -> AnswerStream:
     compiled_graph = load_compiled_graph()
 
-    input = MainInput_a(
-        base_question=config.inputs.search_request.query, log_messages=[]
-    )
+    input = MainInput(log_messages=[])
 
     # Agent search is not a Tool per se, but this is helpful for the frontend
     yield ToolCallKickoff(
@@ -140,7 +142,7 @@ def run_basic_graph(
 ) -> AnswerStream:
     graph = basic_graph_builder()
     compiled_graph = graph.compile()
-    input = BasicInput()
+    input = BasicInput(unused=True)
     return run_graph(compiled_graph, config, input)
 
 
@@ -172,9 +174,7 @@ if __name__ == "__main__":
             # search_request.persona = get_persona_by_id(1, None, db_session)
             # config.perform_initial_search_path_decision = False
             config.behavior.perform_initial_search_decomposition = True
-            input = MainInput_a(
-                base_question=config.inputs.search_request.query, log_messages=[]
-            )
+            input = MainInput(log_messages=[])
 
             tool_responses: list = []
             for output in run_graph(compiled_graph, config, input):

backend/onyx/agents/agent_search/shared_graph_utils/agent_prompt_ops.py

@@ -7,6 +7,7 @@ from onyx.agents.agent_search.models import GraphConfig
 from onyx.agents.agent_search.shared_graph_utils.models import (
     AgentPromptEnrichmentComponents,
 )
+from onyx.agents.agent_search.shared_graph_utils.utils import format_docs
 from onyx.agents.agent_search.shared_graph_utils.utils import (
     get_persona_agent_prompt_expressions,
 )
@@ -40,13 +41,7 @@ def build_sub_question_answer_prompt(
 
     date_str = build_date_time_string()
 
-    # TODO: This should include document metadata and title
-    docs_format_list = [
-        f"Document Number: [D{doc_num + 1}]\nContent: {doc.combined_content}\n\n"
-        for doc_num, doc in enumerate(docs)
-    ]
-
-    docs_str = "\n\n".join(docs_format_list)
+    docs_str = format_docs(docs)
 
     docs_str = trim_prompt_piece(
         config,
@@ -150,3 +145,38 @@ def get_prompt_enrichment_components(
         history=history,
         date_str=date_str,
     )
+
+
+def binary_string_test(text: str, positive_value: str = "yes") -> bool:
+    """
+    Tests if a string contains a positive value (case-insensitive).
+
+    Args:
+        text: The string to test
+        positive_value: The value to look for (defaults to "yes")
+
+    Returns:
+        True if the positive value is found in the text
+    """
+    return positive_value.lower() in text.lower()
+
+
+def binary_string_test_after_answer_separator(
+    text: str, positive_value: str = "yes", separator: str = "Answer:"
+) -> bool:
+    """
+    Tests if a string contains a positive value (case-insensitive).
+
+    Args:
+        text: The string to test
+        positive_value: The value to look for (defaults to "yes")
+
+    Returns:
+        True if the positive value is found in the text
+    """
+
+    if separator not in text:
+        return False
+    relevant_text = text.split(f"{separator}")[-1]
+
+    return binary_string_test(relevant_text, positive_value)

backend/onyx/agents/agent_search/shared_graph_utils/calculations.py

@@ -1,7 +1,11 @@
 import numpy as np
 
+from onyx.agents.agent_search.shared_graph_utils.models import AnswerGenerationDocuments
 from onyx.agents.agent_search.shared_graph_utils.models import RetrievalFitScoreMetrics
 from onyx.agents.agent_search.shared_graph_utils.models import RetrievalFitStats
+from onyx.agents.agent_search.shared_graph_utils.operators import (
+    dedup_inference_section_list,
+)
 from onyx.chat.models import SectionRelevancePiece
 from onyx.context.search.models import InferenceSection
 from onyx.utils.logger import setup_logger
@@ -96,3 +100,106 @@ def get_fit_scores(
     )
 
     return fit_eval
+
+
+def get_answer_generation_documents(
+    relevant_docs: list[InferenceSection],
+    context_documents: list[InferenceSection],
+    original_question_docs: list[InferenceSection],
+    max_docs: int,
+) -> AnswerGenerationDocuments:
+    """
+    Create a deduplicated list of documents to stream, prioritizing relevant docs.
+
+    Args:
+        relevant_docs: Primary documents to include
+        context_documents: Additional context documents to append
+        original_question_docs: Original question documents to append
+        max_docs: Maximum number of documents to return
+
+    Returns:
+        List of deduplicated documents, limited to max_docs
+    """
+    # get relevant_doc ids
+    relevant_doc_ids = [doc.center_chunk.document_id for doc in relevant_docs]
+
+    # Start with relevant docs or fallback to original question docs
+    streaming_documents = relevant_docs.copy()
+
+    # Use a set for O(1) lookups of document IDs
+    seen_doc_ids = {doc.center_chunk.document_id for doc in streaming_documents}
+
+    # Combine additional documents to check in one iteration
+    additional_docs = context_documents + original_question_docs
+    for doc_idx, doc in enumerate(additional_docs):
+        doc_id = doc.center_chunk.document_id
+        if doc_id not in seen_doc_ids:
+            streaming_documents.append(doc)
+            seen_doc_ids.add(doc_id)
+
+    streaming_documents = dedup_inference_section_list(streaming_documents)
+
+    relevant_streaming_docs = [
+        doc
+        for doc in streaming_documents
+        if doc.center_chunk.document_id in relevant_doc_ids
+    ]
+    relevant_streaming_docs = dedup_sort_inference_section_list(relevant_streaming_docs)
+
+    additional_streaming_docs = [
+        doc
+        for doc in streaming_documents
+        if doc.center_chunk.document_id not in relevant_doc_ids
+    ]
+    additional_streaming_docs = dedup_sort_inference_section_list(
+        additional_streaming_docs
+    )
+
+    for doc in additional_streaming_docs:
+        if doc.center_chunk.score:
+            doc.center_chunk.score += -2.0
+        else:
+            doc.center_chunk.score = -2.0
+
+    sorted_streaming_documents = relevant_streaming_docs + additional_streaming_docs
+
+    return AnswerGenerationDocuments(
+        streaming_documents=sorted_streaming_documents[:max_docs],
+        context_documents=relevant_streaming_docs[:max_docs],
+    )
+
+
+def dedup_sort_inference_section_list(
+    sections: list[InferenceSection],
+) -> list[InferenceSection]:
+    """Deduplicates InferenceSections by document_id and sorts by score.
+
+    Args:
+        sections: List of InferenceSections to deduplicate and sort
+
+    Returns:
+        Deduplicated list of InferenceSections sorted by score in descending order
+    """
+    # dedupe/merge with existing framework
+    sections = dedup_inference_section_list(sections)
+
+    # Use dict to deduplicate by document_id, keeping highest scored version
+    unique_sections: dict[str, InferenceSection] = {}
+    for section in sections:
+        doc_id = section.center_chunk.document_id
+        if doc_id not in unique_sections:
+            unique_sections[doc_id] = section
+            continue
+
+        # Keep version with higher score
+        existing_score = unique_sections[doc_id].center_chunk.score or 0
+        new_score = section.center_chunk.score or 0
+        if new_score > existing_score:
+            unique_sections[doc_id] = section
+
+    # Sort by score in descending order, handling None scores
+    sorted_sections = sorted(
+        unique_sections.values(), key=lambda x: x.center_chunk.score or 0, reverse=True
+    )
+
+    return sorted_sections

backend/onyx/agents/agent_search/shared_graph_utils/constants.py

@@ -0,0 +1,19 @@
+from enum import Enum
+
+AGENT_LLM_TIMEOUT_MESSAGE = "The agent timed out. Please try again."
+AGENT_LLM_ERROR_MESSAGE = "The agent encountered an error. Please try again."
+AGENT_LLM_RATELIMIT_MESSAGE = (
+    "The agent encountered a rate limit error. Please try again."
+)
+LLM_ANSWER_ERROR_MESSAGE = "The question was not answered due to an LLM error."
+
+AGENT_POSITIVE_VALUE_STR = "yes"
+AGENT_NEGATIVE_VALUE_STR = "no"
+
+AGENT_ANSWER_SEPARATOR = "Answer:"
+
+
+class AgentLLMErrorType(str, Enum):
+    TIMEOUT = "timeout"
+    RATE_LIMIT = "rate_limit"
+    GENERAL_ERROR = "general_error"

backend/onyx/agents/agent_search/shared_graph_utils/models.py

@@ -1,3 +1,5 @@
+from typing import Any
+
 from pydantic import BaseModel
 
 from onyx.agents.agent_search.deep_search.main.models import (
@@ -56,6 +58,12 @@ class InitialAgentResultStats(BaseModel):
     agent_effectiveness: dict[str, float | int | None]
 
 
+class AgentErrorLog(BaseModel):
+    error_message: str
+    error_type: str
+    error_result: str
+
+
 class RefinedAgentStats(BaseModel):
     revision_doc_efficiency: float | None
     revision_question_efficiency: float | None
@@ -110,6 +118,11 @@ class SubQuestionAnswerResults(BaseModel):
     sub_question_retrieval_stats: AgentChunkRetrievalStats
 
 
+class StructuredSubquestionDocuments(BaseModel):
+    cited_documents: list[InferenceSection]
+    context_documents: list[InferenceSection]
+
+
 class CombinedAgentMetrics(BaseModel):
     timings: AgentTimings
     base_metrics: AgentBaseMetrics | None
@@ -126,3 +139,17 @@ class AgentPromptEnrichmentComponents(BaseModel):
     persona_prompts: PersonaPromptExpressions
     history: str
     date_str: str
+
+
+class LLMNodeErrorStrings(BaseModel):
+    timeout: str = "LLM Timeout Error"
+    rate_limit: str = "LLM Rate Limit Error"
+    general_error: str = "General LLM Error"
+
+
+class AnswerGenerationDocuments(BaseModel):
+    streaming_documents: list[InferenceSection]
+    context_documents: list[InferenceSection]
+
+
+BaseMessage_Content = str | list[str | dict[str, Any]]

backend/onyx/agents/agent_search/shared_graph_utils/operators.py

@@ -12,6 +12,13 @@ def dedup_inference_sections(
     return deduped
 
 
+def dedup_inference_section_list(
+    list: list[InferenceSection],
+) -> list[InferenceSection]:
+    deduped = _merge_sections(list)
+    return deduped
+
+
 def dedup_question_answer_results(
     question_answer_results_1: list[SubQuestionAnswerResults],
     question_answer_results_2: list[SubQuestionAnswerResults],

backend/onyx/agents/agent_search/shared_graph_utils/utils.py

@@ -20,10 +20,18 @@ from onyx.agents.agent_search.models import GraphInputs
 from onyx.agents.agent_search.models import GraphPersistence
 from onyx.agents.agent_search.models import GraphSearchConfig
 from onyx.agents.agent_search.models import GraphTooling
+from onyx.agents.agent_search.shared_graph_utils.models import BaseMessage_Content
 from onyx.agents.agent_search.shared_graph_utils.models import (
     EntityRelationshipTermExtraction,
 )
 from onyx.agents.agent_search.shared_graph_utils.models import PersonaPromptExpressions
+from onyx.agents.agent_search.shared_graph_utils.models import (
+    StructuredSubquestionDocuments,
+)
+from onyx.agents.agent_search.shared_graph_utils.models import SubQuestionAnswerResults
+from onyx.agents.agent_search.shared_graph_utils.operators import (
+    dedup_inference_section_list,
+)
 from onyx.chat.models import AnswerPacket
 from onyx.chat.models import AnswerStyleConfig
 from onyx.chat.models import CitationConfig
@@ -34,6 +42,9 @@ from onyx.chat.models import StreamStopInfo
 from onyx.chat.models import StreamStopReason
 from onyx.chat.models import StreamType
 from onyx.chat.prompt_builder.answer_prompt_builder import AnswerPromptBuilder
+from onyx.configs.agent_configs import (
+    AGENT_TIMEOUT_OVERRIDE_LLM_HISTORY_SUMMARY_GENERATION,
+)
 from onyx.configs.chat_configs import CHAT_TARGET_CHUNK_PERCENTAGE
 from onyx.configs.chat_configs import MAX_CHUNKS_FED_TO_CHAT
 from onyx.configs.constants import DEFAULT_PERSONA_ID
@@ -46,6 +57,8 @@ from onyx.context.search.models import SearchRequest
 from onyx.db.engine import get_session_context_manager
 from onyx.db.persona import get_persona_by_id
 from onyx.db.persona import Persona
+from onyx.llm.chat_llm import LLMRateLimitError
+from onyx.llm.chat_llm import LLMTimeoutError
 from onyx.llm.interfaces import LLM
 from onyx.prompts.agent_search import (
     ASSISTANT_SYSTEM_PROMPT_DEFAULT,
@@ -58,6 +71,7 @@ from onyx.prompts.agent_search import (
 )
 from onyx.prompts.prompt_utils import handle_onyx_date_awareness
 from onyx.tools.force import ForceUseTool
+from onyx.tools.models import SearchToolOverrideKwargs
 from onyx.tools.tool_constructor import SearchToolConfig
 from onyx.tools.tool_implementations.search.search_tool import (
     SEARCH_RESPONSE_SUMMARY_ID,
@@ -65,8 +79,9 @@ from onyx.tools.tool_implementations.search.search_tool import (
 from onyx.tools.tool_implementations.search.search_tool import SearchResponseSummary
 from onyx.tools.tool_implementations.search.search_tool import SearchTool
 from onyx.tools.utils import explicit_tool_calling_supported
+from onyx.utils.logger import setup_logger
 
-BaseMessage_Content = str | list[str | dict[str, Any]]
+logger = setup_logger()
 
 
 # Post-processing
@@ -218,7 +233,10 @@ def get_test_config(
         using_tool_calling_llm=using_tool_calling_llm,
     )
 
-    chat_session_id = os.environ.get("ONYX_AS_CHAT_SESSION_ID")
+    chat_session_id = (
+        os.environ.get("ONYX_AS_CHAT_SESSION_ID")
+        or "00000000-0000-0000-0000-000000000000"
+    )
     assert (
         chat_session_id is not None
     ), "ONYX_AS_CHAT_SESSION_ID must be set for backend tests"
@@ -341,8 +359,12 @@ def retrieve_search_docs(
     with get_session_context_manager() as db_session:
         for tool_response in search_tool.run(
             query=question,
-            force_no_rerank=True,
-            alternate_db_session=db_session,
+            override_kwargs=SearchToolOverrideKwargs(
+                force_no_rerank=True,
+                alternate_db_session=db_session,
+                retrieved_sections_callback=None,
+                skip_query_analysis=False,
+            ),
         ):
             # get retrieved docs to send to the rest of the graph
             if tool_response.id == SEARCH_RESPONSE_SUMMARY_ID:
@@ -372,8 +394,24 @@ def summarize_history(
         )
     )
 
-    history_response = llm.invoke(history_context_prompt)
+    try:
+        history_response = llm.invoke(
+            history_context_prompt,
+            timeout_override=AGENT_TIMEOUT_OVERRIDE_LLM_HISTORY_SUMMARY_GENERATION,
+        )
+    except LLMTimeoutError:
+        logger.error("LLM Timeout Error - summarize history")
+        return (
+            history  # this is what is done at this point anyway, so we default to this
+        )
+    except LLMRateLimitError:
+        logger.error("LLM Rate Limit Error - summarize history")
+        return (
+            history  # this is what is done at this point anyway, so we default to this
+        )
+
     assert isinstance(history_response.content, str)
+
     return history_response.content
 
 
@@ -439,3 +477,27 @@ def remove_document_citations(text: str) -> str:
     #   \d+  - one or more digits
     #   \]   - literal ] character
     return re.sub(r"\[(?:D|Q)?\d+\]", "", text)
+
+
+def get_deduplicated_structured_subquestion_documents(
+    sub_question_results: list[SubQuestionAnswerResults],
+) -> StructuredSubquestionDocuments:
+    """
+    Extract and deduplicate all cited documents from sub-question results.
+
+    Args:
+        sub_question_results: List of sub-question results containing cited documents
+
+    Returns:
+        Deduplicated list of cited documents
+    """
+    cited_docs = [
+        doc for result in sub_question_results for doc in result.cited_documents
+    ]
+    context_docs = [
+        doc for result in sub_question_results for doc in result.context_documents
+    ]
+    return StructuredSubquestionDocuments(
+        cited_documents=dedup_inference_section_list(cited_docs),
+        context_documents=dedup_inference_section_list(context_docs),
+    )

backend/onyx/auth/email_utils.py

@@ -1,7 +1,7 @@
 import smtplib
+from datetime import datetime
 from email.mime.multipart import MIMEMultipart
 from email.mime.text import MIMEText
-from textwrap import dedent
 
 from onyx.configs.app_configs import EMAIL_CONFIGURED
 from onyx.configs.app_configs import EMAIL_FROM
@@ -13,23 +13,150 @@ from onyx.configs.app_configs import WEB_DOMAIN
 from onyx.configs.constants import TENANT_ID_COOKIE_NAME
 from onyx.db.models import User
 
+HTML_EMAIL_TEMPLATE = """\
+<!DOCTYPE html>
+<html lang="en">
+<head>
+  <meta charset="UTF-8">
+  <meta name="viewport" content="width=device-width" />
+  <title>{title}</title>
+  <style>
+    body, table, td, a {{
+      font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, Helvetica, Arial, sans-serif;
+      text-size-adjust: 100%;
+      margin: 0;
+      padding: 0;
+      -webkit-font-smoothing: antialiased;
+      -webkit-text-size-adjust: none;
+    }}
+    body {{
+      background-color: #f7f7f7;
+      color: #333;
+    }}
+    .body-content {{
+      color: #333;
+    }}
+    .email-container {{
+      width: 100%;
+      max-width: 600px;
+      margin: 0 auto;
+      background-color: #ffffff;
+      border-radius: 6px;
+      overflow: hidden;
+      border: 1px solid #eaeaea;
+    }}
+    .header {{
+      background-color: #000000;
+      padding: 20px;
+      text-align: center;
+    }}
+    .header img {{
+      max-width: 140px;
+    }}
+    .body-content {{
+      padding: 20px 30px;
+    }}
+    .title {{
+      font-size: 20px;
+      font-weight: bold;
+      margin: 0 0 10px;
+    }}
+    .message {{
+      font-size: 16px;
+      line-height: 1.5;
+      margin: 0 0 20px;
+    }}
+    .cta-button {{
+      display: inline-block;
+      padding: 12px 20px;
+      background-color: #000000;
+      color: #ffffff !important;
+      text-decoration: none;
+      border-radius: 4px;
+      font-weight: 500;
+    }}
+    .footer {{
+      font-size: 13px;
+      color: #6A7280;
+      text-align: center;
+      padding: 20px;
+    }}
+    .footer a {{
+      color: #6b7280;
+      text-decoration: underline;
+    }}
+  </style>
+</head>
+<body>
+  <table role="presentation" class="email-container" cellpadding="0" cellspacing="0">
+    <tr>
+      <td class="header">
+        <img
+          style="background-color: #ffffff; border-radius: 8px;"
+          src="https://www.onyx.app/logos/customer/onyx.png"
+          alt="Onyx Logo"
+        >
+      </td>
+    </tr>
+    <tr>
+      <td class="body-content">
+        <h1 class="title">{heading}</h1>
+        <div class="message">
+          {message}
+        </div>
+        {cta_block}
+      </td>
+    </tr>
+    <tr>
+      <td class="footer">
+        © {year} Onyx. All rights reserved.
+        <br>
+        Have questions? Join our Slack community <a href="https://join.slack.com/t/onyx-dot-app/shared_invite/zt-2twesxdr6-5iQitKZQpgq~hYIZ~dv3KA">here</a>.
+      </td>
+    </tr>
+  </table>
+</body>
+</html>
+"""
+
+
+def build_html_email(
+    heading: str, message: str, cta_text: str | None = None, cta_link: str | None = None
+) -> str:
+    if cta_text and cta_link:
+        cta_block = f'<a class="cta-button" href="{cta_link}">{cta_text}</a>'
+    else:
+        cta_block = ""
+    return HTML_EMAIL_TEMPLATE.format(
+        title=heading,
+        heading=heading,
+        message=message,
+        cta_block=cta_block,
+        year=datetime.now().year,
+    )
+
 
 def send_email(
     user_email: str,
     subject: str,
-    body: str,
+    html_body: str,
+    text_body: str,
     mail_from: str = EMAIL_FROM,
 ) -> None:
     if not EMAIL_CONFIGURED:
         raise ValueError("Email is not configured.")
 
-    msg = MIMEMultipart()
+    msg = MIMEMultipart("alternative")
     msg["Subject"] = subject
     msg["To"] = user_email
     if mail_from:
         msg["From"] = mail_from
 
-    msg.attach(MIMEText(body))
+    part_text = MIMEText(text_body, "plain")
+    part_html = MIMEText(html_body, "html")
+
+    msg.attach(part_text)
+    msg.attach(part_html)
 
     try:
         with smtplib.SMTP(SMTP_SERVER, SMTP_PORT) as s:
@@ -40,26 +167,44 @@ def send_email(
         raise e
 
 
+def send_subscription_cancellation_email(user_email: str) -> None:
+    # Example usage of the reusable HTML
+    subject = "Your Onyx Subscription Has Been Canceled"
+    heading = "Subscription Canceled"
+    message = (
+        "<p>We’re sorry to see you go.</p>"
+        "<p>Your subscription has been canceled and will end on your next billing date.</p>"
+        "<p>If you change your mind, you can always come back!</p>"
+    )
+    cta_text = "Renew Subscription"
+    cta_link = "https://www.onyx.app/pricing"
+    html_content = build_html_email(heading, message, cta_text, cta_link)
+    text_content = (
+        "We're sorry to see you go.\n"
+        "Your subscription has been canceled and will end on your next billing date.\n"
+        "If you change your mind, visit https://www.onyx.app/pricing"
+    )
+    send_email(user_email, subject, html_content, text_content)
+
+
 def send_user_email_invite(user_email: str, current_user: User) -> None:
     subject = "Invitation to Join Onyx Organization"
-    body = dedent(
-        f"""\
-        Hello,
-
-        You have been invited to join an organization on Onyx.
-
-        To join the organization, please visit the following link:
-
-        {WEB_DOMAIN}/auth/signup?email={user_email}
-
-        You'll be asked to set a password or login with Google to complete your registration.
-
-        Best regards,
-        The Onyx Team
-    """
+    heading = "You've Been Invited!"
+    message = (
+        f"<p>You have been invited by {current_user.email} to join an organization on Onyx.</p>"
+        "<p>To join the organization, please click the button below to set a password "
+        "or login with Google and complete your registration.</p>"
     )
-
-    send_email(user_email, subject, body, current_user.email)
+    cta_text = "Join Organization"
+    cta_link = f"{WEB_DOMAIN}/auth/signup?email={user_email}"
+    html_content = build_html_email(heading, message, cta_text, cta_link)
+    text_content = (
+        f"You have been invited by {current_user.email} to join an organization on Onyx.\n"
+        "To join the organization, please visit the following link:\n"
+        f"{WEB_DOMAIN}/auth/signup?email={user_email}\n"
+        "You'll be asked to set a password or login with Google to complete your registration."
+    )
+    send_email(user_email, subject, html_content, text_content)
 
 
 def send_forgot_password_email(
@@ -68,13 +213,15 @@ def send_forgot_password_email(
     mail_from: str = EMAIL_FROM,
     tenant_id: str | None = None,
 ) -> None:
+    # Builds a forgot password email with or without fancy HTML
     subject = "Onyx Forgot Password"
     link = f"{WEB_DOMAIN}/auth/reset-password?token={token}"
     if tenant_id:
         link += f"&{TENANT_ID_COOKIE_NAME}={tenant_id}"
-        # Keep search param same name as cookie for simplicity
-    body = f"Click the following link to reset your password: {link}"
-    send_email(user_email, subject, body, mail_from)
+    message = f"<p>Click the following link to reset your password:</p><p>{link}</p>"
+    html_content = build_html_email("Reset Your Password", message)
+    text_content = f"Click the following link to reset your password: {link}"
+    send_email(user_email, subject, html_content, text_content, mail_from)
 
 
 def send_user_verification_email(
@@ -82,7 +229,12 @@ def send_user_verification_email(
     token: str,
     mail_from: str = EMAIL_FROM,
 ) -> None:
+    # Builds a verification email
     subject = "Onyx Email Verification"
     link = f"{WEB_DOMAIN}/auth/verify-email?token={token}"
-    body = f"Click the following link to verify your email address: {link}"
-    send_email(user_email, subject, body, mail_from)
+    message = (
+        f"<p>Click the following link to verify your email address:</p><p>{link}</p>"
+    )
+    html_content = build_html_email("Verify Your Email", message)
+    text_content = f"Click the following link to verify your email address: {link}"
+    send_email(user_email, subject, html_content, text_content, mail_from)

backend/onyx/auth/users.py

@@ -94,6 +94,7 @@ from onyx.db.models import User
 from onyx.db.users import get_user_by_email
 from onyx.redis.redis_pool import get_async_redis_connection
 from onyx.redis.redis_pool import get_redis_client
+from onyx.server.utils import BasicAuthenticationError
 from onyx.utils.logger import setup_logger
 from onyx.utils.telemetry import create_milestone_and_report
 from onyx.utils.telemetry import optional_telemetry
@@ -107,11 +108,6 @@ from shared_configs.contextvars import CURRENT_TENANT_ID_CONTEXTVAR
 logger = setup_logger()
 
 
-class BasicAuthenticationError(HTTPException):
-    def __init__(self, detail: str):
-        super().__init__(status_code=status.HTTP_403_FORBIDDEN, detail=detail)
-
-
 def is_user_admin(user: User | None) -> bool:
     if AUTH_TYPE == AuthType.DISABLED:
         return True

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

@@ -1,41 +1,56 @@
 from datetime import timedelta
 from typing import Any
+from typing import cast
 
 from celery import Celery
 from celery import signals
 from celery.beat import PersistentScheduler  # type: ignore
 from celery.signals import beat_init
+from celery.utils.log import get_task_logger
 
 import onyx.background.celery.apps.app_base as app_base
+from onyx.background.celery.tasks.beat_schedule import CLOUD_BEAT_MULTIPLIER_DEFAULT
+from onyx.configs.constants import ONYX_CLOUD_REDIS_RUNTIME
+from onyx.configs.constants import ONYX_CLOUD_TENANT_ID
 from onyx.configs.constants import POSTGRES_CELERY_BEAT_APP_NAME
 from onyx.db.engine import get_all_tenant_ids
 from onyx.db.engine import SqlEngine
-from onyx.utils.logger import setup_logger
+from onyx.redis.redis_pool import get_redis_replica_client
 from onyx.utils.variable_functionality import fetch_versioned_implementation
 from shared_configs.configs import IGNORED_SYNCING_TENANT_LIST
 from shared_configs.configs import MULTI_TENANT
 
-logger = setup_logger(__name__)
+task_logger = get_task_logger(__name__)
 
 celery_app = Celery(__name__)
 celery_app.config_from_object("onyx.background.celery.configs.beat")
 
 
 class DynamicTenantScheduler(PersistentScheduler):
+    """This scheduler is useful because we can dynamically adjust task generation rates
+    through it."""
+
+    RELOAD_INTERVAL = 60
+
     def __init__(self, *args: Any, **kwargs: Any) -> None:
-        logger.info("Initializing DynamicTenantScheduler")
         super().__init__(*args, **kwargs)
-        self._reload_interval = timedelta(minutes=2)
+
+        self.last_beat_multiplier = CLOUD_BEAT_MULTIPLIER_DEFAULT
+
+        self._reload_interval = timedelta(
+            seconds=DynamicTenantScheduler.RELOAD_INTERVAL
+        )
         self._last_reload = self.app.now() - self._reload_interval
+
         # Let the parent class handle store initialization
         self.setup_schedule()
         self._try_updating_schedule()
-        logger.info(f"Set reload interval to {self._reload_interval}")
+        task_logger.info(
+            f"DynamicTenantScheduler initialized: reload_interval={self._reload_interval}"
+        )
 
     def setup_schedule(self) -> None:
-        logger.info("Setting up initial schedule")
         super().setup_schedule()
-        logger.info("Initial schedule setup complete")
 
     def tick(self) -> float:
         retval = super().tick()
@@ -44,36 +59,35 @@ class DynamicTenantScheduler(PersistentScheduler):
             self._last_reload is None
             or (now - self._last_reload) > self._reload_interval
         ):
-            logger.info("Reload interval reached, initiating task update")
+            task_logger.debug("Reload interval reached, initiating task update")
             try:
                 self._try_updating_schedule()
-            except (AttributeError, KeyError) as e:
-                logger.exception(f"Failed to process task configuration: {str(e)}")
-            except Exception as e:
-                logger.exception(f"Unexpected error updating tasks: {str(e)}")
+            except (AttributeError, KeyError):
+                task_logger.exception("Failed to process task configuration")
+            except Exception:
+                task_logger.exception("Unexpected error updating tasks")
 
             self._last_reload = now
-            logger.info("Task update completed, reset reload timer")
+
         return retval
 
     def _generate_schedule(
-        self, tenant_ids: list[str] | list[None]
+        self, tenant_ids: list[str] | list[None], beat_multiplier: float
     ) -> dict[str, dict[str, Any]]:
         """Given a list of tenant id's, generates a new beat schedule for celery."""
-        logger.info("Fetching tasks to schedule")
-
         new_schedule: dict[str, dict[str, Any]] = {}
 
         if MULTI_TENANT:
-            # cloud tasks only need the single task beat across all tenants
+            # cloud tasks are system wide and thus only need to be on the beat schedule
+            # once for all tenants
             get_cloud_tasks_to_schedule = fetch_versioned_implementation(
                 "onyx.background.celery.tasks.beat_schedule",
                 "get_cloud_tasks_to_schedule",
             )
 
-            cloud_tasks_to_schedule: list[
-                dict[str, Any]
-            ] = get_cloud_tasks_to_schedule()
+            cloud_tasks_to_schedule: list[dict[str, Any]] = get_cloud_tasks_to_schedule(
+                beat_multiplier
+            )
             for task in cloud_tasks_to_schedule:
                 task_name = task["name"]
                 cloud_task = {
@@ -82,11 +96,14 @@ class DynamicTenantScheduler(PersistentScheduler):
                     "kwargs": task.get("kwargs", {}),
                 }
                 if options := task.get("options"):
-                    logger.debug(f"Adding options to task {task_name}: {options}")
+                    task_logger.debug(f"Adding options to task {task_name}: {options}")
                     cloud_task["options"] = options
                 new_schedule[task_name] = cloud_task
 
         # regular task beats are multiplied across all tenants
+        # note that currently this just schedules for a single tenant in self hosted
+        # and doesn't do anything in the cloud because it's much more scalable
+        # to schedule a single cloud beat task to dispatch per tenant tasks.
         get_tasks_to_schedule = fetch_versioned_implementation(
             "onyx.background.celery.tasks.beat_schedule", "get_tasks_to_schedule"
         )
@@ -95,7 +112,7 @@ class DynamicTenantScheduler(PersistentScheduler):
 
         for tenant_id in tenant_ids:
             if IGNORED_SYNCING_TENANT_LIST and tenant_id in IGNORED_SYNCING_TENANT_LIST:
-                logger.info(
+                task_logger.debug(
                     f"Skipping tenant {tenant_id} as it is in the ignored syncing list"
                 )
                 continue
@@ -104,14 +121,14 @@ class DynamicTenantScheduler(PersistentScheduler):
                 task_name = task["name"]
                 tenant_task_name = f"{task['name']}-{tenant_id}"
 
-                logger.debug(f"Creating task configuration for {tenant_task_name}")
+                task_logger.debug(f"Creating task configuration for {tenant_task_name}")
                 tenant_task = {
                     "task": task["task"],
                     "schedule": task["schedule"],
                     "kwargs": {"tenant_id": tenant_id},
                 }
                 if options := task.get("options"):
-                    logger.debug(
+                    task_logger.debug(
                         f"Adding options to task {tenant_task_name}: {options}"
                     )
                     tenant_task["options"] = options
@@ -121,44 +138,57 @@ class DynamicTenantScheduler(PersistentScheduler):
 
     def _try_updating_schedule(self) -> None:
         """Only updates the actual beat schedule on the celery app when it changes"""
+        do_update = False
 
-        logger.info("_try_updating_schedule starting")
+        r = get_redis_replica_client(tenant_id=ONYX_CLOUD_TENANT_ID)
+
+        task_logger.debug("_try_updating_schedule starting")
 
         tenant_ids = get_all_tenant_ids()
-        logger.info(f"Found {len(tenant_ids)} IDs")
+        task_logger.debug(f"Found {len(tenant_ids)} IDs")
 
         # get current schedule and extract current tenants
         current_schedule = self.schedule.items()
 
-        # there are no more per tenant beat tasks, so comment this out
-        # NOTE: we may not actualy need this scheduler any more and should
-        # test reverting to a regular beat schedule implementation
+        # get potential new state
+        beat_multiplier = CLOUD_BEAT_MULTIPLIER_DEFAULT
+        beat_multiplier_raw = r.get(f"{ONYX_CLOUD_REDIS_RUNTIME}:beat_multiplier")
+        if beat_multiplier_raw is not None:
+            try:
+                beat_multiplier_bytes = cast(bytes, beat_multiplier_raw)
+                beat_multiplier = float(beat_multiplier_bytes.decode())
+            except ValueError:
+                task_logger.error(
+                    f"Invalid beat_multiplier value: {beat_multiplier_raw}"
+                )
 
-        # current_tenants = set()
-        # for task_name, _ in current_schedule:
-        #     task_name = cast(str, task_name)
-        #     if task_name.startswith(ONYX_CLOUD_CELERY_TASK_PREFIX):
-        #         continue
+        new_schedule = self._generate_schedule(tenant_ids, beat_multiplier)
 
-        #     if "_" in task_name:
-        #         # example: "check-for-condition-tenant_12345678-abcd-efgh-ijkl-12345678"
-        #         # -> "12345678-abcd-efgh-ijkl-12345678"
-        #         current_tenants.add(task_name.split("_")[-1])
-        # logger.info(f"Found {len(current_tenants)} existing items in schedule")
+        # if the schedule or beat multiplier has changed, update
+        while True:
+            if beat_multiplier != self.last_beat_multiplier:
+                do_update = True
+                break
 
-        # for tenant_id in tenant_ids:
-        #     if tenant_id not in current_tenants:
-        #         logger.info(f"Processing new tenant: {tenant_id}")
+            if not DynamicTenantScheduler._compare_schedules(
+                current_schedule, new_schedule
+            ):
+                do_update = True
+                break
 
-        new_schedule = self._generate_schedule(tenant_ids)
+            break
 
-        if DynamicTenantScheduler._compare_schedules(current_schedule, new_schedule):
-            logger.info(
-                "_try_updating_schedule: Current schedule is up to date, no changes needed"
+        if not do_update:
+            # exit early if nothing changed
+            task_logger.info(
+                f"_try_updating_schedule - Schedule unchanged: "
+                f"tasks={len(new_schedule)} "
+                f"beat_multiplier={beat_multiplier}"
             )
             return
 
-        logger.info(
+        # schedule needs updating
+        task_logger.debug(
             "Schedule update required",
             extra={
                 "new_tasks": len(new_schedule),
@@ -185,11 +215,19 @@ class DynamicTenantScheduler(PersistentScheduler):
         # Ensure changes are persisted
         self.sync()
 
-        logger.info("_try_updating_schedule: Schedule updated successfully")
+        task_logger.info(
+            f"_try_updating_schedule - Schedule updated: "
+            f"prev_num_tasks={len(current_schedule)} "
+            f"prev_beat_multiplier={self.last_beat_multiplier} "
+            f"tasks={len(new_schedule)} "
+            f"beat_multiplier={beat_multiplier}"
+        )
+
+        self.last_beat_multiplier = beat_multiplier
 
     @staticmethod
     def _compare_schedules(schedule1: dict, schedule2: dict) -> bool:
-        """Compare schedules to determine if an update is needed.
+        """Compare schedules by task name only to determine if an update is needed.
         True if equivalent, False if not."""
         current_tasks = set(name for name, _ in schedule1)
         new_tasks = set(schedule2.keys())
@@ -201,7 +239,7 @@ class DynamicTenantScheduler(PersistentScheduler):
 
 @beat_init.connect
 def on_beat_init(sender: Any, **kwargs: Any) -> None:
-    logger.info("beat_init signal received.")
+    task_logger.info("beat_init signal received.")
 
     # Celery beat shouldn't touch the db at all. But just setting a low minimum here.
     SqlEngine.set_app_name(POSTGRES_CELERY_BEAT_APP_NAME)

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

@@ -144,7 +144,6 @@ def on_worker_init(sender: Worker, **kwargs: Any) -> None:
     # As currently designed, when this worker starts as "primary", we reinitialize redis
     # to a clean state (for our purposes, anyway)
     r.delete(OnyxRedisLocks.CHECK_VESPA_SYNC_BEAT_LOCK)
-    r.delete(OnyxRedisLocks.MONITOR_VESPA_SYNC_BEAT_LOCK)
 
     r.delete(OnyxRedisConstants.ACTIVE_FENCES)
 

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

@@ -1,3 +1,4 @@
+import copy
 from datetime import timedelta
 from typing import Any
 
@@ -18,7 +19,8 @@ BEAT_EXPIRES_DEFAULT = 15 * 60  # 15 minutes (in seconds)
 
 # hack to slow down task dispatch in the cloud until
 # we have a better implementation (backpressure, etc)
-CLOUD_BEAT_SCHEDULE_MULTIPLIER = 8
+# Note that DynamicTenantScheduler can adjust the runtime value for this via Redis
+CLOUD_BEAT_MULTIPLIER_DEFAULT = 8.0
 
 # tasks that run in either self-hosted on cloud
 beat_task_templates: list[dict] = []
@@ -55,16 +57,7 @@ beat_task_templates.extend(
         {
             "name": "check-for-pruning",
             "task": OnyxCeleryTask.CHECK_FOR_PRUNING,
-            "schedule": timedelta(hours=1),
-            "options": {
-                "priority": OnyxCeleryPriority.MEDIUM,
-                "expires": BEAT_EXPIRES_DEFAULT,
-            },
-        },
-        {
-            "name": "monitor-vespa-sync",
-            "task": OnyxCeleryTask.MONITOR_VESPA_SYNC,
-            "schedule": timedelta(seconds=5),
+            "schedule": timedelta(seconds=20),
             "options": {
                 "priority": OnyxCeleryPriority.MEDIUM,
                 "expires": BEAT_EXPIRES_DEFAULT,
@@ -121,7 +114,7 @@ def make_cloud_generator_task(task: dict[str, Any]) -> dict[str, Any]:
 
     # constant options for cloud beat task generators
     task_schedule: timedelta = task["schedule"]
-    cloud_task["schedule"] = task_schedule * CLOUD_BEAT_SCHEDULE_MULTIPLIER
+    cloud_task["schedule"] = task_schedule
     cloud_task["options"] = {}
     cloud_task["options"]["priority"] = OnyxCeleryPriority.HIGHEST
     cloud_task["options"]["expires"] = BEAT_EXPIRES_DEFAULT
@@ -140,14 +133,14 @@ def make_cloud_generator_task(task: dict[str, Any]) -> dict[str, Any]:
     return cloud_task
 
 
-# tasks that only run in the cloud
-# the name attribute must start with ONYX_CLOUD_CELERY_TASK_PREFIX = "cloud" to be filtered
-# by the DynamicTenantScheduler
-cloud_tasks_to_schedule: list[dict] = [
+# tasks that only run in the cloud and are system wide
+# the name attribute must start with ONYX_CLOUD_CELERY_TASK_PREFIX = "cloud" to be seen
+# by the DynamicTenantScheduler as system wide task and not a per tenant task
+beat_cloud_tasks: list[dict] = [
     # cloud specific tasks
     {
-        "name": f"{ONYX_CLOUD_CELERY_TASK_PREFIX}_check-alembic",
-        "task": OnyxCeleryTask.CLOUD_CHECK_ALEMBIC,
+        "name": f"{ONYX_CLOUD_CELERY_TASK_PREFIX}_monitor-alembic",
+        "task": OnyxCeleryTask.CLOUD_MONITOR_ALEMBIC,
         "schedule": timedelta(hours=1),
         "options": {
             "queue": OnyxCeleryQueues.MONITORING,
@@ -155,20 +148,74 @@ cloud_tasks_to_schedule: list[dict] = [
             "expires": BEAT_EXPIRES_DEFAULT,
         },
     },
+    {
+        "name": f"{ONYX_CLOUD_CELERY_TASK_PREFIX}_monitor-celery-queues",
+        "task": OnyxCeleryTask.CLOUD_MONITOR_CELERY_QUEUES,
+        "schedule": timedelta(seconds=30),
+        "options": {
+            "queue": OnyxCeleryQueues.MONITORING,
+            "priority": OnyxCeleryPriority.HIGH,
+            "expires": BEAT_EXPIRES_DEFAULT,
+        },
+    },
 ]
 
-# generate our cloud and self-hosted beat tasks from the templates
-for beat_task_template in beat_task_templates:
-    cloud_task = make_cloud_generator_task(beat_task_template)
-    cloud_tasks_to_schedule.append(cloud_task)
-
+# tasks that only run self hosted
 tasks_to_schedule: list[dict] = []
 if not MULTI_TENANT:
-    tasks_to_schedule = beat_task_templates
+    tasks_to_schedule.extend(
+        [
+            {
+                "name": "monitor-celery-queues",
+                "task": OnyxCeleryTask.MONITOR_CELERY_QUEUES,
+                "schedule": timedelta(seconds=10),
+                "options": {
+                    "priority": OnyxCeleryPriority.MEDIUM,
+                    "expires": BEAT_EXPIRES_DEFAULT,
+                    "queue": OnyxCeleryQueues.MONITORING,
+                },
+            },
+        ]
+    )
+
+    tasks_to_schedule.extend(beat_task_templates)
 
 
-def get_cloud_tasks_to_schedule() -> list[dict[str, Any]]:
-    return cloud_tasks_to_schedule
+def generate_cloud_tasks(
+    beat_tasks: list[dict], beat_templates: list[dict], beat_multiplier: float
+) -> list[dict[str, Any]]:
+    """
+    beat_tasks: system wide tasks that can be sent as is
+    beat_templates: task templates that will be transformed into per tenant tasks via
+    the cloud_beat_task_generator
+    beat_multiplier: a multiplier that can be applied on top of the task schedule
+    to speed up or slow down the task generation rate. useful in production.
+
+    Returns a list of cloud tasks, which consists of incoming tasks + tasks generated
+    from incoming templates.
+    """
+
+    if beat_multiplier <= 0:
+        raise ValueError("beat_multiplier must be positive!")
+
+    cloud_tasks: list[dict] = []
+
+    # generate our tenant aware cloud tasks from the templates
+    for beat_template in beat_templates:
+        cloud_task = make_cloud_generator_task(beat_template)
+        cloud_tasks.append(cloud_task)
+
+    # factor in the cloud multiplier for the above
+    for cloud_task in cloud_tasks:
+        cloud_task["schedule"] = cloud_task["schedule"] * beat_multiplier
+
+    # add the fixed cloud/system beat tasks. No multiplier for these.
+    cloud_tasks.extend(copy.deepcopy(beat_tasks))
+    return cloud_tasks
+
+
+def get_cloud_tasks_to_schedule(beat_multiplier: float) -> list[dict[str, Any]]:
+    return generate_cloud_tasks(beat_cloud_tasks, beat_task_templates, beat_multiplier)
 
 
 def get_tasks_to_schedule() -> list[dict[str, Any]]:

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

@@ -1,10 +1,14 @@
+import traceback
 from datetime import datetime
 from datetime import timezone
+from typing import Any
+from typing import cast
 
 from celery import Celery
 from celery import shared_task
 from celery import Task
 from celery.exceptions import SoftTimeLimitExceeded
+from redis import Redis
 from redis.lock import Lock as RedisLock
 from sqlalchemy.orm import Session
 
@@ -12,18 +16,35 @@ from onyx.background.celery.apps.app_base import task_logger
 from onyx.configs.app_configs import JOB_TIMEOUT
 from onyx.configs.constants import CELERY_GENERIC_BEAT_LOCK_TIMEOUT
 from onyx.configs.constants import OnyxCeleryTask
+from onyx.configs.constants import OnyxRedisConstants
 from onyx.configs.constants import OnyxRedisLocks
+from onyx.db.connector import fetch_connector_by_id
+from onyx.db.connector_credential_pair import add_deletion_failure_message
+from onyx.db.connector_credential_pair import (
+    delete_connector_credential_pair__no_commit,
+)
 from onyx.db.connector_credential_pair import get_connector_credential_pair_from_id
 from onyx.db.connector_credential_pair import get_connector_credential_pairs
+from onyx.db.document import get_document_ids_for_connector_credential_pair
+from onyx.db.document_set import delete_document_set_cc_pair_relationship__no_commit
 from onyx.db.engine import get_session_with_tenant
 from onyx.db.enums import ConnectorCredentialPairStatus
+from onyx.db.enums import SyncStatus
 from onyx.db.enums import SyncType
+from onyx.db.index_attempt import delete_index_attempts
 from onyx.db.search_settings import get_all_search_settings
 from onyx.db.sync_record import cleanup_sync_records
 from onyx.db.sync_record import insert_sync_record
+from onyx.db.sync_record import update_sync_record_status
 from onyx.redis.redis_connector import RedisConnector
+from onyx.redis.redis_connector_delete import RedisConnectorDelete
 from onyx.redis.redis_connector_delete import RedisConnectorDeletePayload
 from onyx.redis.redis_pool import get_redis_client
+from onyx.redis.redis_pool import get_redis_replica_client
+from onyx.utils.variable_functionality import (
+    fetch_versioned_implementation_with_fallback,
+)
+from onyx.utils.variable_functionality import noop_fallback
 
 
 class TaskDependencyError(RuntimeError):
@@ -42,6 +63,7 @@ def check_for_connector_deletion_task(
     self: Task, *, tenant_id: str | None
 ) -> bool | None:
     r = get_redis_client(tenant_id=tenant_id)
+    r_replica = get_redis_replica_client(tenant_id=tenant_id)
 
     lock_beat: RedisLock = r.lock(
         OnyxRedisLocks.CHECK_CONNECTOR_DELETION_BEAT_LOCK,
@@ -77,6 +99,18 @@ def check_for_connector_deletion_task(
                     # clear the stop signal if it exists ... no longer needed
                     redis_connector.stop.set_fence(False)
 
+        lock_beat.reacquire()
+        keys = cast(set[Any], r_replica.smembers(OnyxRedisConstants.ACTIVE_FENCES))
+        for key in keys:
+            key_bytes = cast(bytes, key)
+
+            if not r.exists(key_bytes):
+                r.srem(OnyxRedisConstants.ACTIVE_FENCES, key_bytes)
+                continue
+
+            key_str = key_bytes.decode("utf-8")
+            if key_str.startswith(RedisConnectorDelete.FENCE_PREFIX):
+                monitor_connector_deletion_taskset(tenant_id, key_bytes, r)
     except SoftTimeLimitExceeded:
         task_logger.info(
             "Soft time limit exceeded, task is being terminated gracefully."
@@ -212,3 +246,158 @@ def try_generate_document_cc_pair_cleanup_tasks(
         redis_connector.delete.set_fence(fence_payload)
 
     return tasks_generated
+
+
+def monitor_connector_deletion_taskset(
+    tenant_id: str | None, key_bytes: bytes, r: Redis
+) -> None:
+    fence_key = key_bytes.decode("utf-8")
+    cc_pair_id_str = RedisConnector.get_id_from_fence_key(fence_key)
+    if cc_pair_id_str is None:
+        task_logger.warning(f"could not parse cc_pair_id from {fence_key}")
+        return
+
+    cc_pair_id = int(cc_pair_id_str)
+
+    redis_connector = RedisConnector(tenant_id, cc_pair_id)
+
+    fence_data = redis_connector.delete.payload
+    if not fence_data:
+        task_logger.warning(
+            f"Connector deletion - fence payload invalid: cc_pair={cc_pair_id}"
+        )
+        return
+
+    if fence_data.num_tasks is None:
+        # the fence is setting up but isn't ready yet
+        return
+
+    remaining = redis_connector.delete.get_remaining()
+    task_logger.info(
+        f"Connector deletion progress: cc_pair={cc_pair_id} remaining={remaining} initial={fence_data.num_tasks}"
+    )
+    if remaining > 0:
+        with get_session_with_tenant(tenant_id) as db_session:
+            update_sync_record_status(
+                db_session=db_session,
+                entity_id=cc_pair_id,
+                sync_type=SyncType.CONNECTOR_DELETION,
+                sync_status=SyncStatus.IN_PROGRESS,
+                num_docs_synced=remaining,
+            )
+        return
+
+    with get_session_with_tenant(tenant_id) as db_session:
+        cc_pair = get_connector_credential_pair_from_id(
+            db_session=db_session,
+            cc_pair_id=cc_pair_id,
+        )
+        if not cc_pair:
+            task_logger.warning(
+                f"Connector deletion - cc_pair not found: cc_pair={cc_pair_id}"
+            )
+            return
+
+        try:
+            doc_ids = get_document_ids_for_connector_credential_pair(
+                db_session, cc_pair.connector_id, cc_pair.credential_id
+            )
+            if len(doc_ids) > 0:
+                # NOTE(rkuo): if this happens, documents somehow got added while
+                # deletion was in progress. Likely a bug gating off pruning and indexing
+                # work before deletion starts.
+                task_logger.warning(
+                    "Connector deletion - documents still found after taskset completion. "
+                    "Clearing the current deletion attempt and allowing deletion to restart: "
+                    f"cc_pair={cc_pair_id} "
+                    f"docs_deleted={fence_data.num_tasks} "
+                    f"docs_remaining={len(doc_ids)}"
+                )
+
+                # We don't want to waive off why we get into this state, but resetting
+                # our attempt and letting the deletion restart is a good way to recover
+                redis_connector.delete.reset()
+                raise RuntimeError(
+                    "Connector deletion - documents still found after taskset completion"
+                )
+
+            # clean up the rest of the related Postgres entities
+            # index attempts
+            delete_index_attempts(
+                db_session=db_session,
+                cc_pair_id=cc_pair_id,
+            )
+
+            # document sets
+            delete_document_set_cc_pair_relationship__no_commit(
+                db_session=db_session,
+                connector_id=cc_pair.connector_id,
+                credential_id=cc_pair.credential_id,
+            )
+
+            # user groups
+            cleanup_user_groups = fetch_versioned_implementation_with_fallback(
+                "onyx.db.user_group",
+                "delete_user_group_cc_pair_relationship__no_commit",
+                noop_fallback,
+            )
+            cleanup_user_groups(
+                cc_pair_id=cc_pair_id,
+                db_session=db_session,
+            )
+
+            # finally, delete the cc-pair
+            delete_connector_credential_pair__no_commit(
+                db_session=db_session,
+                connector_id=cc_pair.connector_id,
+                credential_id=cc_pair.credential_id,
+            )
+            # if there are no credentials left, delete the connector
+            connector = fetch_connector_by_id(
+                db_session=db_session,
+                connector_id=cc_pair.connector_id,
+            )
+            if not connector or not len(connector.credentials):
+                task_logger.info(
+                    "Connector deletion - Found no credentials left for connector, deleting connector"
+                )
+                db_session.delete(connector)
+            db_session.commit()
+
+            update_sync_record_status(
+                db_session=db_session,
+                entity_id=cc_pair_id,
+                sync_type=SyncType.CONNECTOR_DELETION,
+                sync_status=SyncStatus.SUCCESS,
+                num_docs_synced=fence_data.num_tasks,
+            )
+
+        except Exception as e:
+            db_session.rollback()
+            stack_trace = traceback.format_exc()
+            error_message = f"Error: {str(e)}\n\nStack Trace:\n{stack_trace}"
+            add_deletion_failure_message(db_session, cc_pair_id, error_message)
+
+            update_sync_record_status(
+                db_session=db_session,
+                entity_id=cc_pair_id,
+                sync_type=SyncType.CONNECTOR_DELETION,
+                sync_status=SyncStatus.FAILED,
+                num_docs_synced=fence_data.num_tasks,
+            )
+
+            task_logger.exception(
+                f"Connector deletion exceptioned: "
+                f"cc_pair={cc_pair_id} connector={cc_pair.connector_id} credential={cc_pair.credential_id}"
+            )
+            raise e
+
+    task_logger.info(
+        f"Connector deletion succeeded: "
+        f"cc_pair={cc_pair_id} "
+        f"connector={cc_pair.connector_id} "
+        f"credential={cc_pair.credential_id} "
+        f"docs_deleted={fence_data.num_tasks}"
+    )
+
+    redis_connector.delete.reset()

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

@@ -175,6 +175,24 @@ def check_for_doc_permissions_sync(self: Task, *, tenant_id: str | None) -> bool
                 )
 
             r.set(OnyxRedisSignals.BLOCK_VALIDATE_PERMISSION_SYNC_FENCES, 1, ex=300)
+
+        # use a lookup table to find active fences. We still have to verify the fence
+        # exists since it is an optimization and not the source of truth.
+        lock_beat.reacquire()
+        keys = cast(set[Any], r_replica.smembers(OnyxRedisConstants.ACTIVE_FENCES))
+        for key in keys:
+            key_bytes = cast(bytes, key)
+
+            if not r.exists(key_bytes):
+                r.srem(OnyxRedisConstants.ACTIVE_FENCES, key_bytes)
+                continue
+
+            key_str = key_bytes.decode("utf-8")
+            if key_str.startswith(RedisConnectorPermissionSync.FENCE_PREFIX):
+                with get_session_with_tenant(tenant_id) as db_session:
+                    monitor_ccpair_permissions_taskset(
+                        tenant_id, key_bytes, r, db_session
+                    )
     except SoftTimeLimitExceeded:
         task_logger.info(
             "Soft time limit exceeded, task is being terminated gracefully."
@@ -349,6 +367,7 @@ def connector_permission_sync_generator_task(
         OnyxRedisLocks.CONNECTOR_DOC_PERMISSIONS_SYNC_LOCK_PREFIX
         + f"_{redis_connector.id}",
         timeout=CELERY_PERMISSIONS_SYNC_LOCK_TIMEOUT,
+        thread_local=False,
     )
 
     acquired = lock.acquire(blocking=False)
@@ -459,14 +478,15 @@ def update_external_document_permissions_task(
     )
     doc_id = document_external_access.doc_id
     external_access = document_external_access.external_access
+
     try:
         with get_session_with_tenant(tenant_id) as db_session:
-            # Add the users to the DB if they don't exist
             batch_add_ext_perm_user_if_not_exists(
                 db_session=db_session,
                 emails=list(external_access.external_user_emails),
+                continue_on_error=True,
             )
-            # Then we upsert the document's external permissions in postgres
+            # Then upsert the document's external permissions
             created_new_doc = upsert_document_external_perms(
                 db_session=db_session,
                 doc_id=doc_id,
@@ -490,11 +510,11 @@ def update_external_document_permissions_task(
                 f"action=update_permissions "
                 f"elapsed={elapsed:.2f}"
             )
+
     except Exception:
         task_logger.exception(
             f"Exception in update_external_document_permissions_task: "
-            f"connector_id={connector_id} "
-            f"doc_id={doc_id}"
+            f"connector_id={connector_id} doc_id={doc_id}"
         )
         return False
 
@@ -755,7 +775,7 @@ class PermissionSyncCallback(IndexingHeartbeatInterface):
             raise
 
 
-"""Monitoring CCPair permissions utils, called in monitor_vespa_sync"""
+"""Monitoring CCPair permissions utils"""
 
 
 def monitor_ccpair_permissions_taskset(

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

@@ -26,11 +26,11 @@ from ee.onyx.external_permissions.sync_params import (
 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_unacked_task_ids
+from onyx.background.error_logging import emit_background_error
 from onyx.configs.app_configs import JOB_TIMEOUT
 from onyx.configs.constants import CELERY_EXTERNAL_GROUP_SYNC_LOCK_TIMEOUT
 from onyx.configs.constants import CELERY_GENERIC_BEAT_LOCK_TIMEOUT
 from onyx.configs.constants import CELERY_TASK_WAIT_FOR_FENCE_TIMEOUT
-from onyx.configs.constants import DANSWER_REDIS_FUNCTION_LOCK_PREFIX
 from onyx.configs.constants import OnyxCeleryPriority
 from onyx.configs.constants import OnyxCeleryQueues
 from onyx.configs.constants import OnyxCeleryTask
@@ -72,18 +72,26 @@ def _is_external_group_sync_due(cc_pair: ConnectorCredentialPair) -> bool:
     """Returns boolean indicating if external group sync is due."""
 
     if cc_pair.access_type != AccessType.SYNC:
-        return False
-
-    # skip external group sync if not active
-    if cc_pair.status != ConnectorCredentialPairStatus.ACTIVE:
+        task_logger.error(
+            f"Recieved non-sync CC Pair {cc_pair.id} for external "
+            f"group sync. Actual access type: {cc_pair.access_type}"
+        )
         return False
 
     if cc_pair.status == ConnectorCredentialPairStatus.DELETING:
+        task_logger.debug(
+            f"Skipping group sync for CC Pair {cc_pair.id} - "
+            f"CC Pair is being deleted"
+        )
         return False
 
     # If there is not group sync function for the connector, we don't run the sync
     # This is fine because all sources dont necessarily have a concept of groups
     if not GROUP_PERMISSIONS_FUNC_MAP.get(cc_pair.connector.source):
+        task_logger.debug(
+            f"Skipping group sync for CC Pair {cc_pair.id} - "
+            f"no group sync function for {cc_pair.connector.source}"
+        )
         return False
 
     # If the last sync is None, it has never been run so we run the sync
@@ -125,6 +133,9 @@ def check_for_external_group_sync(self: Task, *, tenant_id: str | None) -> bool
 
     # these tasks should never overlap
     if not lock_beat.acquire(blocking=False):
+        task_logger.warning(
+            f"Failed to acquire beat lock for external group sync: {tenant_id}"
+        )
         return None
 
     try:
@@ -205,20 +216,12 @@ def try_creating_external_group_sync_task(
 
     redis_connector = RedisConnector(tenant_id, cc_pair_id)
 
-    LOCK_TIMEOUT = 30
-
-    lock: RedisLock = r.lock(
-        DANSWER_REDIS_FUNCTION_LOCK_PREFIX + "try_generate_external_group_sync_tasks",
-        timeout=LOCK_TIMEOUT,
-    )
-
-    acquired = lock.acquire(blocking_timeout=LOCK_TIMEOUT / 2)
-    if not acquired:
-        return None
-
     try:
         # Dont kick off a new sync if the previous one is still running
         if redis_connector.external_group_sync.fenced:
+            logger.warning(
+                f"Skipping external group sync for CC Pair {cc_pair_id} - already running."
+            )
             return None
 
         redis_connector.external_group_sync.generator_clear()
@@ -269,9 +272,6 @@ def try_creating_external_group_sync_task(
             f"Unexpected exception while trying to create external group sync task: cc_pair={cc_pair_id}"
         )
         return None
-    finally:
-        if lock.owned():
-            lock.release()
 
     return payload_id
 
@@ -304,22 +304,26 @@ def connector_external_group_sync_generator_task(
     start = time.monotonic()
     while True:
         if time.monotonic() - start > CELERY_TASK_WAIT_FOR_FENCE_TIMEOUT:
-            raise ValueError(
+            msg = (
                 f"connector_external_group_sync_generator_task - timed out waiting for fence to be ready: "
                 f"fence={redis_connector.external_group_sync.fence_key}"
             )
+            emit_background_error(msg, cc_pair_id=cc_pair_id)
+            raise ValueError(msg)
 
         if not redis_connector.external_group_sync.fenced:  # The fence must exist
-            raise ValueError(
+            msg = (
                 f"connector_external_group_sync_generator_task - fence not found: "
                 f"fence={redis_connector.external_group_sync.fence_key}"
             )
+            emit_background_error(msg, cc_pair_id=cc_pair_id)
+            raise ValueError(msg)
 
         payload = redis_connector.external_group_sync.payload  # The payload must exist
         if not payload:
-            raise ValueError(
-                "connector_external_group_sync_generator_task: payload invalid or not found"
-            )
+            msg = "connector_external_group_sync_generator_task: payload invalid or not found"
+            emit_background_error(msg, cc_pair_id=cc_pair_id)
+            raise ValueError(msg)
 
         if payload.celery_task_id is None:
             logger.info(
@@ -344,9 +348,9 @@ def connector_external_group_sync_generator_task(
 
     acquired = lock.acquire(blocking=False)
     if not acquired:
-        task_logger.warning(
-            f"External group sync task already running, exiting...: cc_pair={cc_pair_id}"
-        )
+        msg = f"External group sync task already running, exiting...: cc_pair={cc_pair_id}"
+        emit_background_error(msg, cc_pair_id=cc_pair_id)
+        task_logger.error(msg)
         return None
 
     try:
@@ -367,9 +371,9 @@ def connector_external_group_sync_generator_task(
 
             ext_group_sync_func = GROUP_PERMISSIONS_FUNC_MAP.get(source_type)
             if ext_group_sync_func is None:
-                raise ValueError(
-                    f"No external group sync func found for {source_type} for cc_pair: {cc_pair_id}"
-                )
+                msg = f"No external group sync func found for {source_type} for cc_pair: {cc_pair_id}"
+                emit_background_error(msg, cc_pair_id=cc_pair_id)
+                raise ValueError(msg)
 
             logger.info(
                 f"Syncing external groups for {source_type} for cc_pair: {cc_pair_id}"
@@ -400,9 +404,9 @@ def connector_external_group_sync_generator_task(
                 sync_status=SyncStatus.SUCCESS,
             )
     except Exception as e:
-        task_logger.exception(
-            f"External group sync exceptioned: cc_pair={cc_pair_id} payload_id={payload.id}"
-        )
+        msg = f"External group sync exceptioned: cc_pair={cc_pair_id} payload_id={payload.id}"
+        task_logger.exception(msg)
+        emit_background_error(msg + f"\n\n{e}", cc_pair_id=cc_pair_id)
 
         with get_session_with_tenant(tenant_id) as db_session:
             update_sync_record_status(
@@ -492,9 +496,11 @@ def validate_external_group_sync_fence(
     fence_key = key_bytes.decode("utf-8")
     cc_pair_id_str = RedisConnector.get_id_from_fence_key(fence_key)
     if cc_pair_id_str is None:
-        task_logger.warning(
+        msg = (
             f"validate_external_group_sync_fence - could not parse id from {fence_key}"
         )
+        emit_background_error(msg)
+        task_logger.error(msg)
         return
 
     cc_pair_id = int(cc_pair_id_str)
@@ -509,12 +515,14 @@ def validate_external_group_sync_fence(
     try:
         payload = redis_connector.external_group_sync.payload
     except ValidationError:
-        task_logger.exception(
+        msg = (
             "validate_external_group_sync_fence - "
             "Resetting fence because fence schema is out of date: "
             f"cc_pair={cc_pair_id} "
             f"fence={fence_key}"
         )
+        task_logger.exception(msg)
+        emit_background_error(msg, cc_pair_id=cc_pair_id)
 
         redis_connector.external_group_sync.reset()
         return
@@ -551,12 +559,15 @@ def validate_external_group_sync_fence(
     # return
 
     # celery tasks don't exist and the active signal has expired, possibly due to a crash. Clean it up.
-    logger.warning(
-        "validate_external_group_sync_fence - "
-        "Resetting fence because no associated celery tasks were found: "
-        f"cc_pair={cc_pair_id} "
-        f"fence={fence_key} "
-        f"payload_id={payload.id}"
+    emit_background_error(
+        message=(
+            "validate_external_group_sync_fence - "
+            "Resetting fence because no associated celery tasks were found: "
+            f"cc_pair={cc_pair_id} "
+            f"fence={fence_key} "
+            f"payload_id={payload.id}"
+        ),
+        cc_pair_id=cc_pair_id,
     )
 
     redis_connector.external_group_sync.reset()

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

@@ -6,13 +6,18 @@ from datetime import datetime
 from datetime import timezone
 from http import HTTPStatus
 from time import sleep
+from typing import Any
+from typing import cast
 
 import sentry_sdk
 from celery import shared_task
 from celery import Task
 from celery.exceptions import SoftTimeLimitExceeded
+from celery.result import AsyncResult
+from celery.states import READY_STATES
 from redis import Redis
 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_utils import httpx_init_vespa_pool
@@ -30,6 +35,7 @@ from onyx.configs.constants import CELERY_GENERIC_BEAT_LOCK_TIMEOUT
 from onyx.configs.constants import CELERY_INDEXING_LOCK_TIMEOUT
 from onyx.configs.constants import CELERY_TASK_WAIT_FOR_FENCE_TIMEOUT
 from onyx.configs.constants import OnyxCeleryTask
+from onyx.configs.constants import OnyxRedisConstants
 from onyx.configs.constants import OnyxRedisLocks
 from onyx.configs.constants import OnyxRedisSignals
 from onyx.db.connector import mark_ccpair_with_indexing_trigger
@@ -37,6 +43,7 @@ from onyx.db.connector_credential_pair import fetch_connector_credential_pairs
 from onyx.db.connector_credential_pair import get_connector_credential_pair_from_id
 from onyx.db.engine import get_session_with_tenant
 from onyx.db.enums import IndexingMode
+from onyx.db.enums import IndexingStatus
 from onyx.db.index_attempt import get_index_attempt
 from onyx.db.index_attempt import get_last_attempt_for_cc_pair
 from onyx.db.index_attempt import mark_attempt_canceled
@@ -47,9 +54,12 @@ from onyx.db.swap_index import check_index_swap
 from onyx.natural_language_processing.search_nlp_models import EmbeddingModel
 from onyx.natural_language_processing.search_nlp_models import warm_up_bi_encoder
 from onyx.redis.redis_connector import RedisConnector
+from onyx.redis.redis_connector_index import RedisConnectorIndex
 from onyx.redis.redis_pool import get_redis_client
 from onyx.redis.redis_pool import get_redis_replica_client
 from onyx.redis.redis_pool import redis_lock_dump
+from onyx.redis.redis_pool import SCAN_ITER_COUNT_DEFAULT
+from onyx.redis.redis_utils import is_fence
 from onyx.utils.logger import setup_logger
 from onyx.utils.variable_functionality import global_version
 from shared_configs.configs import INDEXING_MODEL_SERVER_HOST
@@ -60,6 +70,150 @@ from shared_configs.configs import SENTRY_DSN
 logger = setup_logger()
 
 
+def monitor_ccpair_indexing_taskset(
+    tenant_id: str | None, key_bytes: bytes, r: Redis, db_session: Session
+) -> None:
+    # if the fence doesn't exist, there's nothing to do
+    fence_key = key_bytes.decode("utf-8")
+    composite_id = RedisConnector.get_id_from_fence_key(fence_key)
+    if composite_id is None:
+        task_logger.warning(
+            f"Connector indexing: could not parse composite_id from {fence_key}"
+        )
+        return
+
+    # parse out metadata and initialize the helper class with it
+    parts = composite_id.split("/")
+    if len(parts) != 2:
+        return
+
+    cc_pair_id = int(parts[0])
+    search_settings_id = int(parts[1])
+
+    redis_connector = RedisConnector(tenant_id, cc_pair_id)
+    redis_connector_index = redis_connector.new_index(search_settings_id)
+    if not redis_connector_index.fenced:
+        return
+
+    payload = redis_connector_index.payload
+    if not payload:
+        return
+
+    elapsed_started_str = None
+    if payload.started:
+        elapsed_started = datetime.now(timezone.utc) - payload.started
+        elapsed_started_str = f"{elapsed_started.total_seconds():.2f}"
+
+    elapsed_submitted = datetime.now(timezone.utc) - payload.submitted
+
+    progress = redis_connector_index.get_progress()
+    if progress is not None:
+        task_logger.info(
+            f"Connector indexing progress: "
+            f"attempt={payload.index_attempt_id} "
+            f"cc_pair={cc_pair_id} "
+            f"search_settings={search_settings_id} "
+            f"progress={progress} "
+            f"elapsed_submitted={elapsed_submitted.total_seconds():.2f} "
+            f"elapsed_started={elapsed_started_str}"
+        )
+
+    if payload.index_attempt_id is None or payload.celery_task_id is None:
+        # the task is still setting up
+        return
+
+    # never use any blocking methods on the result from inside a task!
+    result: AsyncResult = AsyncResult(payload.celery_task_id)
+
+    # inner/outer/inner double check pattern to avoid race conditions when checking for
+    # bad state
+
+    # Verify: if the generator isn't complete, the task must not be in READY state
+    # inner = get_completion / generator_complete not signaled
+    # outer = result.state in READY state
+    status_int = redis_connector_index.get_completion()
+    if status_int is None:  # inner signal not set ... possible error
+        task_state = result.state
+        if (
+            task_state in READY_STATES
+        ):  # outer signal in terminal state ... possible error
+            # Now double check!
+            if redis_connector_index.get_completion() is None:
+                # inner signal still not set (and cannot change when outer result_state is READY)
+                # Task is finished but generator complete isn't set.
+                # We have a problem! Worker may have crashed.
+                task_result = str(result.result)
+                task_traceback = str(result.traceback)
+
+                msg = (
+                    f"Connector indexing aborted or exceptioned: "
+                    f"attempt={payload.index_attempt_id} "
+                    f"celery_task={payload.celery_task_id} "
+                    f"cc_pair={cc_pair_id} "
+                    f"search_settings={search_settings_id} "
+                    f"elapsed_submitted={elapsed_submitted.total_seconds():.2f} "
+                    f"result.state={task_state} "
+                    f"result.result={task_result} "
+                    f"result.traceback={task_traceback}"
+                )
+                task_logger.warning(msg)
+
+                try:
+                    index_attempt = get_index_attempt(
+                        db_session, payload.index_attempt_id
+                    )
+                    if index_attempt:
+                        if (
+                            index_attempt.status != IndexingStatus.CANCELED
+                            and index_attempt.status != IndexingStatus.FAILED
+                        ):
+                            mark_attempt_failed(
+                                index_attempt_id=payload.index_attempt_id,
+                                db_session=db_session,
+                                failure_reason=msg,
+                            )
+                except Exception:
+                    task_logger.exception(
+                        "Connector indexing - Transient exception marking index attempt as failed: "
+                        f"attempt={payload.index_attempt_id} "
+                        f"tenant={tenant_id} "
+                        f"cc_pair={cc_pair_id} "
+                        f"search_settings={search_settings_id}"
+                    )
+
+                redis_connector_index.reset()
+        return
+
+    if redis_connector_index.watchdog_signaled():
+        # if the generator is complete, don't clean up until the watchdog has exited
+        task_logger.info(
+            f"Connector indexing - Delaying finalization until watchdog has exited: "
+            f"attempt={payload.index_attempt_id} "
+            f"cc_pair={cc_pair_id} "
+            f"search_settings={search_settings_id} "
+            f"progress={progress} "
+            f"elapsed_submitted={elapsed_submitted.total_seconds():.2f} "
+            f"elapsed_started={elapsed_started_str}"
+        )
+
+        return
+
+    status_enum = HTTPStatus(status_int)
+
+    task_logger.info(
+        f"Connector indexing finished: "
+        f"attempt={payload.index_attempt_id} "
+        f"cc_pair={cc_pair_id} "
+        f"search_settings={search_settings_id} "
+        f"progress={progress} "
+        f"status={status_enum.name} "
+        f"elapsed_submitted={elapsed_submitted.total_seconds():.2f} "
+        f"elapsed_started={elapsed_started_str}"
+    )
+
+    redis_connector_index.reset()
+
+
 @shared_task(
     name=OnyxCeleryTask.CHECK_FOR_INDEXING,
     soft_time_limit=300,
@@ -91,6 +245,25 @@ def check_for_indexing(self: Task, *, tenant_id: str | None) -> int | None:
     try:
         locked = True
 
+        # SPECIAL 0/3: sync lookup table for active fences
+        # we want to run this less frequently than the overall task
+        if not redis_client.exists(OnyxRedisSignals.BLOCK_BUILD_FENCE_LOOKUP_TABLE):
+            # build a lookup table of existing fences
+            # this is just a migration concern and should be unnecessary once
+            # lookup tables are rolled out
+            for key_bytes in redis_client_replica.scan_iter(
+                count=SCAN_ITER_COUNT_DEFAULT
+            ):
+                if is_fence(key_bytes) and not redis_client.sismember(
+                    OnyxRedisConstants.ACTIVE_FENCES, key_bytes
+                ):
+                    logger.warning(f"Adding {key_bytes} to the lookup table.")
+                    redis_client.sadd(OnyxRedisConstants.ACTIVE_FENCES, key_bytes)
+
+            redis_client.set(OnyxRedisSignals.BLOCK_BUILD_FENCE_LOOKUP_TABLE, 1, ex=300)
+
+        # 1/3: KICKOFF
+
         # check for search settings swap
         with get_session_with_tenant(tenant_id=tenant_id) as db_session:
             old_search_settings = check_index_swap(db_session=db_session)
@@ -197,6 +370,8 @@ def check_for_indexing(self: Task, *, tenant_id: str | None) -> int | None:
 
         lock_beat.reacquire()
 
+        # 2/3: VALIDATE
+
         # Fail any index attempts in the DB that don't have fences
         # This shouldn't ever happen!
         with get_session_with_tenant(tenant_id) as db_session:
@@ -236,6 +411,26 @@ def check_for_indexing(self: Task, *, tenant_id: str | None) -> int | None:
                 task_logger.exception("Exception while validating indexing fences")
 
             redis_client.set(OnyxRedisSignals.BLOCK_VALIDATE_INDEXING_FENCES, 1, ex=60)
+
+        # 3/3: FINALIZE
+        lock_beat.reacquire()
+        keys = cast(
+            set[Any], redis_client_replica.smembers(OnyxRedisConstants.ACTIVE_FENCES)
+        )
+        for key in keys:
+            key_bytes = cast(bytes, key)
+
+            if not redis_client.exists(key_bytes):
+                redis_client.srem(OnyxRedisConstants.ACTIVE_FENCES, key_bytes)
+                continue
+
+            key_str = key_bytes.decode("utf-8")
+            if key_str.startswith(RedisConnectorIndex.FENCE_PREFIX):
+                with get_session_with_tenant(tenant_id) as db_session:
+                    monitor_ccpair_indexing_taskset(
+                        tenant_id, key_bytes, redis_client_replica, db_session
+                    )
+
     except SoftTimeLimitExceeded:
         task_logger.info(
             "Soft time limit exceeded, task is being terminated gracefully."

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

@@ -17,7 +17,8 @@ from sqlalchemy import text
 from sqlalchemy.orm import Session
 
 from onyx.background.celery.apps.app_base import task_logger
-from onyx.background.celery.tasks.vespa.tasks import celery_get_queue_length
+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.configs.constants import CELERY_GENERIC_BEAT_LOCK_TIMEOUT
 from onyx.configs.constants import ONYX_CLOUD_TENANT_ID
 from onyx.configs.constants import OnyxCeleryQueues
@@ -189,9 +190,9 @@ def _build_connector_start_latency_metric(
         desired_start_time = cc_pair.connector.time_created
     else:
         if not cc_pair.connector.refresh_freq:
-            task_logger.error(
-                "Found non-initial index attempt for connector "
-                "without refresh_freq. This should never happen."
+            task_logger.debug(
+                "Connector has no refresh_freq and this is a non-initial index attempt. "
+                "Assuming user manually triggered indexing, so we'll skip start latency metric."
             )
             return None
 
@@ -420,6 +421,7 @@ def _collect_sync_metrics(db_session: Session, redis_std: Redis) -> list[Metric]
       - Throughput (docs/min) (only if success)
       - Raw start/end times for each sync
     """
+
     one_hour_ago = get_db_current_time(db_session) - timedelta(hours=1)
 
     # Get all sync records that ended in the last hour
@@ -587,6 +589,10 @@ def _collect_sync_metrics(db_session: Session, redis_std: Redis) -> list[Metric]
                 entity = db_session.scalar(
                     select(UserGroup).where(UserGroup.id == sync_record.entity_id)
                 )
+            else:
+                # Only user groups and document set sync records have
+                #  an associated entity we can use for latency metrics
+                continue
 
             if entity is None:
                 task_logger.error(
@@ -717,7 +723,7 @@ def monitor_background_processes(self: Task, *, tenant_id: str | None) -> None:
 
 
 @shared_task(
-    name=OnyxCeleryTask.CLOUD_CHECK_ALEMBIC,
+    name=OnyxCeleryTask.CLOUD_MONITOR_ALEMBIC,
 )
 def cloud_check_alembic() -> bool | None:
     """A task to verify that all tenants are on the same alembic revision.
@@ -777,7 +783,7 @@ def cloud_check_alembic() -> bool | None:
 
                     tenant_to_revision[tenant_id] = result_scalar
                 except Exception:
-                    task_logger.warning(f"Tenant {tenant_id} has no revision!")
+                    task_logger.error(f"Tenant {tenant_id} has no revision!")
                     tenant_to_revision[tenant_id] = ALEMBIC_NULL_REVISION
 
         # get the total count of each revision
@@ -847,3 +853,55 @@ def cloud_check_alembic() -> bool | None:
         f"cloud_check_alembic finished: num_tenants={len(tenant_ids)} elapsed={time_elapsed:.2f}"
     )
     return True
+
+
+@shared_task(
+    name=OnyxCeleryTask.CLOUD_MONITOR_CELERY_QUEUES, ignore_result=True, bind=True
+)
+def cloud_monitor_celery_queues(
+    self: Task,
+) -> None:
+    return monitor_celery_queues_helper(self)
+
+
+@shared_task(name=OnyxCeleryTask.MONITOR_CELERY_QUEUES, ignore_result=True, bind=True)
+def monitor_celery_queues(self: Task, *, tenant_id: str | None) -> None:
+    return monitor_celery_queues_helper(self)
+
+
+def monitor_celery_queues_helper(
+    task: Task,
+) -> None:
+    """A task to monitor all celery queue lengths."""
+
+    r_celery = task.app.broker_connection().channel().client  # type: ignore
+    n_celery = celery_get_queue_length("celery", r_celery)
+    n_indexing = celery_get_queue_length(OnyxCeleryQueues.CONNECTOR_INDEXING, r_celery)
+    n_sync = celery_get_queue_length(OnyxCeleryQueues.VESPA_METADATA_SYNC, r_celery)
+    n_deletion = celery_get_queue_length(OnyxCeleryQueues.CONNECTOR_DELETION, r_celery)
+    n_pruning = celery_get_queue_length(OnyxCeleryQueues.CONNECTOR_PRUNING, r_celery)
+    n_permissions_sync = celery_get_queue_length(
+        OnyxCeleryQueues.CONNECTOR_DOC_PERMISSIONS_SYNC, r_celery
+    )
+    n_external_group_sync = celery_get_queue_length(
+        OnyxCeleryQueues.CONNECTOR_EXTERNAL_GROUP_SYNC, r_celery
+    )
+    n_permissions_upsert = celery_get_queue_length(
+        OnyxCeleryQueues.DOC_PERMISSIONS_UPSERT, r_celery
+    )
+
+    n_indexing_prefetched = celery_get_unacked_task_ids(
+        OnyxCeleryQueues.CONNECTOR_INDEXING, r_celery
+    )
+
+    task_logger.info(
+        f"Queue lengths: celery={n_celery} "
+        f"indexing={n_indexing} "
+        f"indexing_prefetched={len(n_indexing_prefetched)} "
+        f"sync={n_sync} "
+        f"deletion={n_deletion} "
+        f"pruning={n_pruning} "
+        f"permissions_sync={n_permissions_sync} "
+        f"external_group_sync={n_external_group_sync} "
+        f"permissions_upsert={n_permissions_upsert} "
+    )

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

@@ -122,34 +122,39 @@ def check_for_pruning(self: Task, *, tenant_id: str | None) -> bool | None:
         return None
 
     try:
-        cc_pair_ids: list[int] = []
-        with get_session_with_tenant(tenant_id) as db_session:
-            cc_pairs = get_connector_credential_pairs(db_session)
-            for cc_pair_entry in cc_pairs:
-                cc_pair_ids.append(cc_pair_entry.id)
+        # the entire task needs to run frequently in order to finalize pruning
 
-        for cc_pair_id in cc_pair_ids:
-            lock_beat.reacquire()
+        # but pruning only kicks off once per hour
+        if not r.exists(OnyxRedisSignals.BLOCK_PRUNING):
+            cc_pair_ids: list[int] = []
             with get_session_with_tenant(tenant_id) as db_session:
-                cc_pair = get_connector_credential_pair_from_id(
-                    db_session=db_session,
-                    cc_pair_id=cc_pair_id,
-                )
-                if not cc_pair:
-                    continue
+                cc_pairs = get_connector_credential_pairs(db_session)
+                for cc_pair_entry in cc_pairs:
+                    cc_pair_ids.append(cc_pair_entry.id)
 
-                if not _is_pruning_due(cc_pair):
-                    continue
+            for cc_pair_id in cc_pair_ids:
+                lock_beat.reacquire()
+                with get_session_with_tenant(tenant_id) as db_session:
+                    cc_pair = get_connector_credential_pair_from_id(
+                        db_session=db_session,
+                        cc_pair_id=cc_pair_id,
+                    )
+                    if not cc_pair:
+                        continue
 
-                payload_id = try_creating_prune_generator_task(
-                    self.app, cc_pair, db_session, r, tenant_id
-                )
-                if not payload_id:
-                    continue
+                    if not _is_pruning_due(cc_pair):
+                        continue
 
-                task_logger.info(
-                    f"Pruning queued: cc_pair={cc_pair.id} id={payload_id}"
-                )
+                    payload_id = try_creating_prune_generator_task(
+                        self.app, cc_pair, db_session, r, tenant_id
+                    )
+                    if not payload_id:
+                        continue
+
+                    task_logger.info(
+                        f"Pruning queued: cc_pair={cc_pair.id} id={payload_id}"
+                    )
+            r.set(OnyxRedisSignals.BLOCK_PRUNING, 1, ex=3600)
 
         # we want to run this less frequently than the overall task
         lock_beat.reacquire()
@@ -163,6 +168,22 @@ def check_for_pruning(self: Task, *, tenant_id: str | None) -> bool | None:
                 task_logger.exception("Exception while validating pruning fences")
 
             r.set(OnyxRedisSignals.BLOCK_VALIDATE_PRUNING_FENCES, 1, ex=300)
+
+        # use a lookup table to find active fences. We still have to verify the fence
+        # exists since it is an optimization and not the source of truth.
+        lock_beat.reacquire()
+        keys = cast(set[Any], r_replica.smembers(OnyxRedisConstants.ACTIVE_FENCES))
+        for key in keys:
+            key_bytes = cast(bytes, key)
+
+            if not r.exists(key_bytes):
+                r.srem(OnyxRedisConstants.ACTIVE_FENCES, key_bytes)
+                continue
+
+            key_str = key_bytes.decode("utf-8")
+            if key_str.startswith(RedisConnectorPrune.FENCE_PREFIX):
+                with get_session_with_tenant(tenant_id) as db_session:
+                    monitor_ccpair_pruning_taskset(tenant_id, key_bytes, r, db_session)
     except SoftTimeLimitExceeded:
         task_logger.info(
             "Soft time limit exceeded, task is being terminated gracefully."
@@ -481,7 +502,7 @@ def connector_pruning_generator_task(
     )
 
 
-"""Monitoring pruning utils, called in monitor_vespa_sync"""
+"""Monitoring pruning utils"""
 
 
 def monitor_ccpair_pruning_taskset(

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

@@ -8,6 +8,7 @@ from celery.exceptions import SoftTimeLimitExceeded
 from redis.lock import Lock as RedisLock
 from tenacity import RetryError
 
+from ee.onyx.server.tenants.product_gating import get_gated_tenants
 from onyx.access.access import get_access_for_document
 from onyx.background.celery.apps.app_base import task_logger
 from onyx.background.celery.tasks.beat_schedule import BEAT_EXPIRES_DEFAULT
@@ -252,7 +253,11 @@ def cloud_beat_task_generator(
 
     try:
         tenant_ids = get_all_tenant_ids()
+        gated_tenants = get_gated_tenants()
         for tenant_id in tenant_ids:
+            if tenant_id in gated_tenants:
+                continue
+
             current_time = time.monotonic()
             if current_time - last_lock_time >= (CELERY_GENERIC_BEAT_LOCK_TIMEOUT / 4):
                 lock_beat.reacquire()
@@ -270,6 +275,7 @@ def cloud_beat_task_generator(
                 queue=queue,
                 priority=priority,
                 expires=expires,
+                ignore_result=True,
             )
     except SoftTimeLimitExceeded:
         task_logger.info(

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

@@ -1,9 +1,5 @@
-import random
 import time
-import traceback
 from collections.abc import Callable
-from datetime import datetime
-from datetime import timezone
 from http import HTTPStatus
 from typing import Any
 from typing import cast
@@ -13,8 +9,6 @@ from celery import Celery
 from celery import shared_task
 from celery import Task
 from celery.exceptions import SoftTimeLimitExceeded
-from celery.result import AsyncResult
-from celery.states import READY_STATES
 from redis import Redis
 from redis.lock import Lock as RedisLock
 from sqlalchemy.orm import Session
@@ -22,47 +16,27 @@ from tenacity import RetryError
 
 from onyx.access.access import get_access_for_document
 from onyx.background.celery.apps.app_base import task_logger
-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.tasks.doc_permission_syncing.tasks import (
-    monitor_ccpair_permissions_taskset,
-)
-from onyx.background.celery.tasks.pruning.tasks import monitor_ccpair_pruning_taskset
 from onyx.background.celery.tasks.shared.RetryDocumentIndex import RetryDocumentIndex
 from onyx.background.celery.tasks.shared.tasks import LIGHT_SOFT_TIME_LIMIT
 from onyx.background.celery.tasks.shared.tasks import LIGHT_TIME_LIMIT
 from onyx.configs.app_configs import JOB_TIMEOUT
 from onyx.configs.app_configs import VESPA_SYNC_MAX_TASKS
 from onyx.configs.constants import CELERY_VESPA_SYNC_BEAT_LOCK_TIMEOUT
-from onyx.configs.constants import OnyxCeleryQueues
 from onyx.configs.constants import OnyxCeleryTask
 from onyx.configs.constants import OnyxRedisConstants
 from onyx.configs.constants import OnyxRedisLocks
-from onyx.configs.constants import OnyxRedisSignals
-from onyx.db.connector import fetch_connector_by_id
-from onyx.db.connector_credential_pair import add_deletion_failure_message
-from onyx.db.connector_credential_pair import (
-    delete_connector_credential_pair__no_commit,
-)
-from onyx.db.connector_credential_pair import get_connector_credential_pair_from_id
 from onyx.db.connector_credential_pair import get_connector_credential_pairs
 from onyx.db.document import count_documents_by_needs_sync
 from onyx.db.document import get_document
-from onyx.db.document import get_document_ids_for_connector_credential_pair
 from onyx.db.document import mark_document_as_synced
 from onyx.db.document_set import delete_document_set
-from onyx.db.document_set import delete_document_set_cc_pair_relationship__no_commit
 from onyx.db.document_set import fetch_document_sets
 from onyx.db.document_set import fetch_document_sets_for_document
 from onyx.db.document_set import get_document_set_by_id
 from onyx.db.document_set import mark_document_set_as_synced
 from onyx.db.engine import get_session_with_tenant
-from onyx.db.enums import IndexingStatus
 from onyx.db.enums import SyncStatus
 from onyx.db.enums import SyncType
-from onyx.db.index_attempt import delete_index_attempts
-from onyx.db.index_attempt import get_index_attempt
-from onyx.db.index_attempt import mark_attempt_failed
 from onyx.db.models import DocumentSet
 from onyx.db.models import UserGroup
 from onyx.db.search_settings import get_active_search_settings
@@ -72,20 +46,14 @@ from onyx.db.sync_record import update_sync_record_status
 from onyx.document_index.factory import get_default_document_index
 from onyx.document_index.interfaces import VespaDocumentFields
 from onyx.httpx.httpx_pool import HttpxPool
-from onyx.redis.redis_connector import RedisConnector
 from onyx.redis.redis_connector_credential_pair import RedisConnectorCredentialPair
 from onyx.redis.redis_connector_credential_pair import (
     RedisGlobalConnectorCredentialPair,
 )
-from onyx.redis.redis_connector_delete import RedisConnectorDelete
-from onyx.redis.redis_connector_doc_perm_sync import RedisConnectorPermissionSync
-from onyx.redis.redis_connector_index import RedisConnectorIndex
-from onyx.redis.redis_connector_prune import RedisConnectorPrune
 from onyx.redis.redis_document_set import RedisDocumentSet
 from onyx.redis.redis_pool import get_redis_client
 from onyx.redis.redis_pool import get_redis_replica_client
 from onyx.redis.redis_pool import redis_lock_dump
-from onyx.redis.redis_pool import SCAN_ITER_COUNT_DEFAULT
 from onyx.redis.redis_usergroup import RedisUserGroup
 from onyx.utils.logger import setup_logger
 from onyx.utils.variable_functionality import fetch_versioned_implementation
@@ -94,7 +62,6 @@ from onyx.utils.variable_functionality import (
 )
 from onyx.utils.variable_functionality import global_version
 from onyx.utils.variable_functionality import noop_fallback
-from shared_configs.configs import MULTI_TENANT
 
 logger = setup_logger()
 
@@ -111,9 +78,14 @@ logger = setup_logger()
 def check_for_vespa_sync_task(self: Task, *, tenant_id: str | None) -> bool | None:
     """Runs periodically to check if any document needs syncing.
     Generates sets of tasks for Celery if syncing is needed."""
+
+    # Useful for debugging timing issues with reacquisitions. TODO: remove once more generalized logging is in place
+    task_logger.info("check_for_vespa_sync_task started")
+
     time_start = time.monotonic()
 
     r = get_redis_client(tenant_id=tenant_id)
+    r_replica = get_redis_replica_client(tenant_id=tenant_id)
 
     lock_beat: RedisLock = r.lock(
         OnyxRedisLocks.CHECK_VESPA_SYNC_BEAT_LOCK,
@@ -125,6 +97,7 @@ def check_for_vespa_sync_task(self: Task, *, tenant_id: str | None) -> bool | No
         return None
 
     try:
+        # 1/3: KICKOFF
         with get_session_with_tenant(tenant_id) as db_session:
             try_generate_stale_document_sync_tasks(
                 self.app, VESPA_SYNC_MAX_TASKS, db_session, r, lock_beat, tenant_id
@@ -151,9 +124,8 @@ def check_for_vespa_sync_task(self: Task, *, tenant_id: str | None) -> bool | No
         # endregion
 
         # check if any user groups are not synced
+        lock_beat.reacquire()
         if global_version.is_ee_version():
-            lock_beat.reacquire()
-
             try:
                 fetch_user_groups = fetch_versioned_implementation(
                     "onyx.db.user_group", "fetch_user_groups"
@@ -179,6 +151,35 @@ def check_for_vespa_sync_task(self: Task, *, tenant_id: str | None) -> bool | No
                             self.app, usergroup_id, db_session, r, lock_beat, tenant_id
                         )
 
+        # 2/3: VALIDATE: TODO
+
+        # 3/3: FINALIZE
+        lock_beat.reacquire()
+        keys = cast(set[Any], r_replica.smembers(OnyxRedisConstants.ACTIVE_FENCES))
+        for key in keys:
+            key_bytes = cast(bytes, key)
+
+            if not r.exists(key_bytes):
+                r.srem(OnyxRedisConstants.ACTIVE_FENCES, key_bytes)
+                continue
+
+            key_str = key_bytes.decode("utf-8")
+            if key_str == RedisGlobalConnectorCredentialPair.FENCE_KEY:
+                monitor_connector_taskset(r)
+            elif key_str.startswith(RedisDocumentSet.FENCE_PREFIX):
+                with get_session_with_tenant(tenant_id) as db_session:
+                    monitor_document_set_taskset(tenant_id, key_bytes, r, db_session)
+            elif key_str.startswith(RedisUserGroup.FENCE_PREFIX):
+                monitor_usergroup_taskset = (
+                    fetch_versioned_implementation_with_fallback(
+                        "onyx.background.celery.tasks.vespa.tasks",
+                        "monitor_usergroup_taskset",
+                        noop_fallback,
+                    )
+                )
+                with get_session_with_tenant(tenant_id) as db_session:
+                    monitor_usergroup_taskset(tenant_id, key_bytes, r, db_session)
+
     except SoftTimeLimitExceeded:
         task_logger.info(
             "Soft time limit exceeded, task is being terminated gracefully."
@@ -495,484 +496,23 @@ def monitor_document_set_taskset(
             task_logger.info(
                 f"Successfully synced document set: document_set={document_set_id}"
             )
-        update_sync_record_status(
-            db_session=db_session,
-            entity_id=document_set_id,
-            sync_type=SyncType.DOCUMENT_SET,
-            sync_status=SyncStatus.SUCCESS,
-            num_docs_synced=initial_count,
-        )
-
-    rds.reset()
-
-
-def monitor_connector_deletion_taskset(
-    tenant_id: str | None, key_bytes: bytes, r: Redis
-) -> None:
-    fence_key = key_bytes.decode("utf-8")
-    cc_pair_id_str = RedisConnector.get_id_from_fence_key(fence_key)
-    if cc_pair_id_str is None:
-        task_logger.warning(f"could not parse cc_pair_id from {fence_key}")
-        return
-
-    cc_pair_id = int(cc_pair_id_str)
-
-    redis_connector = RedisConnector(tenant_id, cc_pair_id)
-
-    fence_data = redis_connector.delete.payload
-    if not fence_data:
-        task_logger.warning(
-            f"Connector deletion - fence payload invalid: cc_pair={cc_pair_id}"
-        )
-        return
-
-    if fence_data.num_tasks is None:
-        # the fence is setting up but isn't ready yet
-        return
-
-    remaining = redis_connector.delete.get_remaining()
-    task_logger.info(
-        f"Connector deletion progress: cc_pair={cc_pair_id} remaining={remaining} initial={fence_data.num_tasks}"
-    )
-    if remaining > 0:
-        with get_session_with_tenant(tenant_id) as db_session:
-            update_sync_record_status(
-                db_session=db_session,
-                entity_id=cc_pair_id,
-                sync_type=SyncType.CONNECTOR_DELETION,
-                sync_status=SyncStatus.IN_PROGRESS,
-                num_docs_synced=remaining,
-            )
-        return
-
-    with get_session_with_tenant(tenant_id) as db_session:
-        cc_pair = get_connector_credential_pair_from_id(
-            db_session=db_session,
-            cc_pair_id=cc_pair_id,
-        )
-        if not cc_pair:
-            task_logger.warning(
-                f"Connector deletion - cc_pair not found: cc_pair={cc_pair_id}"
-            )
-            return
 
         try:
-            doc_ids = get_document_ids_for_connector_credential_pair(
-                db_session, cc_pair.connector_id, cc_pair.credential_id
-            )
-            if len(doc_ids) > 0:
-                # NOTE(rkuo): if this happens, documents somehow got added while
-                # deletion was in progress. Likely a bug gating off pruning and indexing
-                # work before deletion starts.
-                task_logger.warning(
-                    "Connector deletion - documents still found after taskset completion. "
-                    "Clearing the current deletion attempt and allowing deletion to restart: "
-                    f"cc_pair={cc_pair_id} "
-                    f"docs_deleted={fence_data.num_tasks} "
-                    f"docs_remaining={len(doc_ids)}"
-                )
-
-                # We don't want to waive off why we get into this state, but resetting
-                # our attempt and letting the deletion restart is a good way to recover
-                redis_connector.delete.reset()
-                raise RuntimeError(
-                    "Connector deletion - documents still found after taskset completion"
-                )
-
-            # clean up the rest of the related Postgres entities
-            # index attempts
-            delete_index_attempts(
-                db_session=db_session,
-                cc_pair_id=cc_pair_id,
-            )
-
-            # document sets
-            delete_document_set_cc_pair_relationship__no_commit(
-                db_session=db_session,
-                connector_id=cc_pair.connector_id,
-                credential_id=cc_pair.credential_id,
-            )
-
-            # user groups
-            cleanup_user_groups = fetch_versioned_implementation_with_fallback(
-                "onyx.db.user_group",
-                "delete_user_group_cc_pair_relationship__no_commit",
-                noop_fallback,
-            )
-            cleanup_user_groups(
-                cc_pair_id=cc_pair_id,
-                db_session=db_session,
-            )
-
-            # finally, delete the cc-pair
-            delete_connector_credential_pair__no_commit(
-                db_session=db_session,
-                connector_id=cc_pair.connector_id,
-                credential_id=cc_pair.credential_id,
-            )
-            # if there are no credentials left, delete the connector
-            connector = fetch_connector_by_id(
-                db_session=db_session,
-                connector_id=cc_pair.connector_id,
-            )
-            if not connector or not len(connector.credentials):
-                task_logger.info(
-                    "Connector deletion - Found no credentials left for connector, deleting connector"
-                )
-                db_session.delete(connector)
-            db_session.commit()
-
             update_sync_record_status(
                 db_session=db_session,
-                entity_id=cc_pair_id,
-                sync_type=SyncType.CONNECTOR_DELETION,
+                entity_id=document_set_id,
+                sync_type=SyncType.DOCUMENT_SET,
                 sync_status=SyncStatus.SUCCESS,
-                num_docs_synced=fence_data.num_tasks,
+                num_docs_synced=initial_count,
             )
-
-        except Exception as e:
-            db_session.rollback()
-            stack_trace = traceback.format_exc()
-            error_message = f"Error: {str(e)}\n\nStack Trace:\n{stack_trace}"
-            add_deletion_failure_message(db_session, cc_pair_id, error_message)
-
-            update_sync_record_status(
-                db_session=db_session,
-                entity_id=cc_pair_id,
-                sync_type=SyncType.CONNECTOR_DELETION,
-                sync_status=SyncStatus.FAILED,
-                num_docs_synced=fence_data.num_tasks,
-            )
-
+        except Exception:
             task_logger.exception(
-                f"Connector deletion exceptioned: "
-                f"cc_pair={cc_pair_id} connector={cc_pair.connector_id} credential={cc_pair.credential_id}"
-            )
-            raise e
-
-    task_logger.info(
-        f"Connector deletion succeeded: "
-        f"cc_pair={cc_pair_id} "
-        f"connector={cc_pair.connector_id} "
-        f"credential={cc_pair.credential_id} "
-        f"docs_deleted={fence_data.num_tasks}"
-    )
-
-    redis_connector.delete.reset()
-
-
-def monitor_ccpair_indexing_taskset(
-    tenant_id: str | None, key_bytes: bytes, r: Redis, db_session: Session
-) -> None:
-    # if the fence doesn't exist, there's nothing to do
-    fence_key = key_bytes.decode("utf-8")
-    composite_id = RedisConnector.get_id_from_fence_key(fence_key)
-    if composite_id is None:
-        task_logger.warning(
-            f"Connector indexing: could not parse composite_id from {fence_key}"
-        )
-        return
-
-    # parse out metadata and initialize the helper class with it
-    parts = composite_id.split("/")
-    if len(parts) != 2:
-        return
-
-    cc_pair_id = int(parts[0])
-    search_settings_id = int(parts[1])
-
-    redis_connector = RedisConnector(tenant_id, cc_pair_id)
-    redis_connector_index = redis_connector.new_index(search_settings_id)
-    if not redis_connector_index.fenced:
-        return
-
-    payload = redis_connector_index.payload
-    if not payload:
-        return
-
-    elapsed_started_str = None
-    if payload.started:
-        elapsed_started = datetime.now(timezone.utc) - payload.started
-        elapsed_started_str = f"{elapsed_started.total_seconds():.2f}"
-
-    elapsed_submitted = datetime.now(timezone.utc) - payload.submitted
-
-    progress = redis_connector_index.get_progress()
-    if progress is not None:
-        task_logger.info(
-            f"Connector indexing progress: "
-            f"attempt={payload.index_attempt_id} "
-            f"cc_pair={cc_pair_id} "
-            f"search_settings={search_settings_id} "
-            f"progress={progress} "
-            f"elapsed_submitted={elapsed_submitted.total_seconds():.2f} "
-            f"elapsed_started={elapsed_started_str}"
-        )
-
-    if payload.index_attempt_id is None or payload.celery_task_id is None:
-        # the task is still setting up
-        return
-
-    # never use any blocking methods on the result from inside a task!
-    result: AsyncResult = AsyncResult(payload.celery_task_id)
-
-    # inner/outer/inner double check pattern to avoid race conditions when checking for
-    # bad state
-
-    # Verify: if the generator isn't complete, the task must not be in READY state
-    # inner = get_completion / generator_complete not signaled
-    # outer = result.state in READY state
-    status_int = redis_connector_index.get_completion()
-    if status_int is None:  # inner signal not set ... possible error
-        task_state = result.state
-        if (
-            task_state in READY_STATES
-        ):  # outer signal in terminal state ... possible error
-            # Now double check!
-            if redis_connector_index.get_completion() is None:
-                # inner signal still not set (and cannot change when outer result_state is READY)
-                # Task is finished but generator complete isn't set.
-                # We have a problem! Worker may have crashed.
-                task_result = str(result.result)
-                task_traceback = str(result.traceback)
-
-                msg = (
-                    f"Connector indexing aborted or exceptioned: "
-                    f"attempt={payload.index_attempt_id} "
-                    f"celery_task={payload.celery_task_id} "
-                    f"cc_pair={cc_pair_id} "
-                    f"search_settings={search_settings_id} "
-                    f"elapsed_submitted={elapsed_submitted.total_seconds():.2f} "
-                    f"result.state={task_state} "
-                    f"result.result={task_result} "
-                    f"result.traceback={task_traceback}"
-                )
-                task_logger.warning(msg)
-
-                try:
-                    index_attempt = get_index_attempt(
-                        db_session, payload.index_attempt_id
-                    )
-                    if index_attempt:
-                        if (
-                            index_attempt.status != IndexingStatus.CANCELED
-                            and index_attempt.status != IndexingStatus.FAILED
-                        ):
-                            mark_attempt_failed(
-                                index_attempt_id=payload.index_attempt_id,
-                                db_session=db_session,
-                                failure_reason=msg,
-                            )
-                except Exception:
-                    task_logger.exception(
-                        "Connector indexing - Transient exception marking index attempt as failed: "
-                        f"attempt={payload.index_attempt_id} "
-                        f"tenant={tenant_id} "
-                        f"cc_pair={cc_pair_id} "
-                        f"search_settings={search_settings_id}"
-                    )
-
-                redis_connector_index.reset()
-        return
-
-    if redis_connector_index.watchdog_signaled():
-        # if the generator is complete, don't clean up until the watchdog has exited
-        task_logger.info(
-            f"Connector indexing - Delaying finalization until watchdog has exited: "
-            f"attempt={payload.index_attempt_id} "
-            f"cc_pair={cc_pair_id} "
-            f"search_settings={search_settings_id} "
-            f"progress={progress} "
-            f"elapsed_submitted={elapsed_submitted.total_seconds():.2f} "
-            f"elapsed_started={elapsed_started_str}"
-        )
-
-        return
-
-    status_enum = HTTPStatus(status_int)
-
-    task_logger.info(
-        f"Connector indexing finished: "
-        f"attempt={payload.index_attempt_id} "
-        f"cc_pair={cc_pair_id} "
-        f"search_settings={search_settings_id} "
-        f"progress={progress} "
-        f"status={status_enum.name} "
-        f"elapsed_submitted={elapsed_submitted.total_seconds():.2f} "
-        f"elapsed_started={elapsed_started_str}"
-    )
-
-    redis_connector_index.reset()
-
-
-@shared_task(
-    name=OnyxCeleryTask.MONITOR_VESPA_SYNC,
-    ignore_result=True,
-    soft_time_limit=300,
-    bind=True,
-)
-def monitor_vespa_sync(self: Task, tenant_id: str | None) -> bool | None:
-    """This is a celery beat task that monitors and finalizes various long running tasks.
-
-    The name monitor_vespa_sync is a bit of a misnomer since it checks many different tasks
-    now. Should change that at some point.
-
-    It scans for fence values and then gets the counts of any associated tasksets.
-    For many tasks, the count is 0, that means all tasks finished and we should clean up.
-
-    This task lock timeout is CELERY_METADATA_SYNC_BEAT_LOCK_TIMEOUT seconds, so don't
-    do anything too expensive in this function!
-
-    Returns True if the task actually did work, False if it exited early to prevent overlap
-    """
-    task_logger.info(f"monitor_vespa_sync starting: tenant={tenant_id}")
-
-    time_start = time.monotonic()
-
-    r = get_redis_client(tenant_id=tenant_id)
-
-    # Replica usage notes
-    #
-    # False negatives are OK. (aka fail to to see a key that exists on the master).
-    # We simply skip the monitoring work and it will be caught on the next pass.
-    #
-    # False positives are not OK, and are possible if we clear a fence on the master and
-    # then read from the replica. In this case, monitoring work could be done on a fence
-    # that no longer exists. To avoid this, we scan from the replica, but double check
-    # the result on the master.
-    r_replica = get_redis_replica_client(tenant_id=tenant_id)
-
-    lock_beat: RedisLock = r.lock(
-        OnyxRedisLocks.MONITOR_VESPA_SYNC_BEAT_LOCK,
-        timeout=CELERY_VESPA_SYNC_BEAT_LOCK_TIMEOUT,
-    )
-
-    # prevent overlapping tasks
-    if not lock_beat.acquire(blocking=False):
-        return None
-
-    try:
-        # print current queue lengths
-        time.monotonic()
-        # we don't need every tenant polling redis for this info.
-        if not MULTI_TENANT or random.randint(1, 10) == 10:
-            r_celery = self.app.broker_connection().channel().client  # type: ignore
-            n_celery = celery_get_queue_length("celery", r_celery)
-            n_indexing = celery_get_queue_length(
-                OnyxCeleryQueues.CONNECTOR_INDEXING, r_celery
-            )
-            n_sync = celery_get_queue_length(
-                OnyxCeleryQueues.VESPA_METADATA_SYNC, r_celery
-            )
-            n_deletion = celery_get_queue_length(
-                OnyxCeleryQueues.CONNECTOR_DELETION, r_celery
-            )
-            n_pruning = celery_get_queue_length(
-                OnyxCeleryQueues.CONNECTOR_PRUNING, r_celery
-            )
-            n_permissions_sync = celery_get_queue_length(
-                OnyxCeleryQueues.CONNECTOR_DOC_PERMISSIONS_SYNC, r_celery
-            )
-            n_external_group_sync = celery_get_queue_length(
-                OnyxCeleryQueues.CONNECTOR_EXTERNAL_GROUP_SYNC, r_celery
-            )
-            n_permissions_upsert = celery_get_queue_length(
-                OnyxCeleryQueues.DOC_PERMISSIONS_UPSERT, r_celery
+                "update_sync_record_status exceptioned. "
+                f"document_set_id={document_set_id} "
+                "Resetting document set regardless."
             )
 
-            prefetched = celery_get_unacked_task_ids(
-                OnyxCeleryQueues.CONNECTOR_INDEXING, r_celery
-            )
-
-            task_logger.info(
-                f"Queue lengths: celery={n_celery} "
-                f"indexing={n_indexing} "
-                f"indexing_prefetched={len(prefetched)} "
-                f"sync={n_sync} "
-                f"deletion={n_deletion} "
-                f"pruning={n_pruning} "
-                f"permissions_sync={n_permissions_sync} "
-                f"external_group_sync={n_external_group_sync} "
-                f"permissions_upsert={n_permissions_upsert} "
-            )
-
-        # we want to run this less frequently than the overall task
-        if not r.exists(OnyxRedisSignals.BLOCK_BUILD_FENCE_LOOKUP_TABLE):
-            # build a lookup table of existing fences
-            # this is just a migration concern and should be unnecessary once
-            # lookup tables are rolled out
-            for key_bytes in r_replica.scan_iter(count=SCAN_ITER_COUNT_DEFAULT):
-                if is_fence(key_bytes) and not r.sismember(
-                    OnyxRedisConstants.ACTIVE_FENCES, key_bytes
-                ):
-                    logger.warning(f"Adding {key_bytes} to the lookup table.")
-                    r.sadd(OnyxRedisConstants.ACTIVE_FENCES, key_bytes)
-
-            r.set(OnyxRedisSignals.BLOCK_BUILD_FENCE_LOOKUP_TABLE, 1, ex=300)
-
-        # use a lookup table to find active fences. We still have to verify the fence
-        # exists since it is an optimization and not the source of truth.
-        keys = cast(set[Any], r_replica.smembers(OnyxRedisConstants.ACTIVE_FENCES))
-        for key in keys:
-            key_bytes = cast(bytes, key)
-
-            if not r.exists(key_bytes):
-                r.srem(OnyxRedisConstants.ACTIVE_FENCES, key_bytes)
-                continue
-
-            key_str = key_bytes.decode("utf-8")
-            if key_str == RedisGlobalConnectorCredentialPair.FENCE_KEY:
-                monitor_connector_taskset(r)
-            elif key_str.startswith(RedisDocumentSet.FENCE_PREFIX):
-                with get_session_with_tenant(tenant_id) as db_session:
-                    monitor_document_set_taskset(tenant_id, key_bytes, r, db_session)
-            elif key_str.startswith(RedisUserGroup.FENCE_PREFIX):
-                monitor_usergroup_taskset = (
-                    fetch_versioned_implementation_with_fallback(
-                        "onyx.background.celery.tasks.vespa.tasks",
-                        "monitor_usergroup_taskset",
-                        noop_fallback,
-                    )
-                )
-                with get_session_with_tenant(tenant_id) as db_session:
-                    monitor_usergroup_taskset(tenant_id, key_bytes, r, db_session)
-            elif key_str.startswith(RedisConnectorDelete.FENCE_PREFIX):
-                monitor_connector_deletion_taskset(tenant_id, key_bytes, r)
-            elif key_str.startswith(RedisConnectorPrune.FENCE_PREFIX):
-                with get_session_with_tenant(tenant_id) as db_session:
-                    monitor_ccpair_pruning_taskset(tenant_id, key_bytes, r, db_session)
-            elif key_str.startswith(RedisConnectorIndex.FENCE_PREFIX):
-                with get_session_with_tenant(tenant_id) as db_session:
-                    monitor_ccpair_indexing_taskset(tenant_id, key_bytes, r, db_session)
-            elif key_str.startswith(RedisConnectorPermissionSync.FENCE_PREFIX):
-                with get_session_with_tenant(tenant_id) as db_session:
-                    monitor_ccpair_permissions_taskset(
-                        tenant_id, key_bytes, r, db_session
-                    )
-            else:
-                pass
-    except SoftTimeLimitExceeded:
-        task_logger.info(
-            "Soft time limit exceeded, task is being terminated gracefully."
-        )
-        return False
-    except Exception:
-        task_logger.exception("monitor_vespa_sync exceptioned.")
-        return False
-    finally:
-        if lock_beat.owned():
-            lock_beat.release()
-        else:
-            task_logger.error(
-                "monitor_vespa_sync - Lock not owned on completion: "
-                f"tenant={tenant_id}"
-                # f"timings={timings}"
-            )
-            redis_lock_dump(lock_beat, r)
-
-    time_elapsed = time.monotonic() - time_start
-    task_logger.info(f"monitor_vespa_sync finished: elapsed={time_elapsed:.2f}")
-    return True
+    rds.reset()
 
 
 @shared_task(
@@ -1072,23 +612,3 @@ def vespa_metadata_sync_task(
         self.retry(exc=e, countdown=countdown)
 
     return True
-
-
-def is_fence(key_bytes: bytes) -> bool:
-    key_str = key_bytes.decode("utf-8")
-    if key_str == RedisGlobalConnectorCredentialPair.FENCE_KEY:
-        return True
-    if key_str.startswith(RedisDocumentSet.FENCE_PREFIX):
-        return True
-    if key_str.startswith(RedisUserGroup.FENCE_PREFIX):
-        return True
-    if key_str.startswith(RedisConnectorDelete.FENCE_PREFIX):
-        return True
-    if key_str.startswith(RedisConnectorPrune.FENCE_PREFIX):
-        return True
-    if key_str.startswith(RedisConnectorIndex.FENCE_PREFIX):
-        return True
-    if key_str.startswith(RedisConnectorPermissionSync.FENCE_PREFIX):
-        return True
-
-    return False

backend/onyx/background/error_logging.py

@@ -0,0 +1,13 @@
+from onyx.db.background_error import create_background_error
+from onyx.db.engine import get_session_with_tenant
+
+
+def emit_background_error(
+    message: str,
+    cc_pair_id: int | None = None,
+) -> None:
+    """Currently just saves a row in the background_errors table.
+
+    In the future, could create notifications based on the severity."""
+    with get_session_with_tenant() as db_session:
+        create_background_error(db_session, message, cc_pair_id)

backend/onyx/configs/agent_configs.py

@@ -3,6 +3,24 @@ import os
 INITIAL_SEARCH_DECOMPOSITION_ENABLED = True
 ALLOW_REFINEMENT = True
 
+AGENT_DEFAULT_RETRIEVAL_HITS = 15
+AGENT_DEFAULT_RERANKING_HITS = 10
+AGENT_DEFAULT_SUB_QUESTION_MAX_CONTEXT_HITS = 8
+AGENT_DEFAULT_NUM_DOCS_FOR_INITIAL_DECOMPOSITION = 3
+AGENT_DEFAULT_NUM_DOCS_FOR_REFINED_DECOMPOSITION = 5
+
+AGENT_DEFAULT_MAX_STREAMED_DOCS_FOR_INITIAL_ANSWER = 25
+AGENT_DEFAULT_MAX_STREAMED_DOCS_FOR_REFINED_ANSWER = 35
+
+
+AGENT_DEFAULT_EXPLORATORY_SEARCH_RESULTS = 5
+AGENT_DEFAULT_MIN_ORIG_QUESTION_DOCS = 3
+AGENT_DEFAULT_MAX_ANSWER_CONTEXT_DOCS = 10
+AGENT_DEFAULT_MAX_STATIC_HISTORY_WORD_LENGTH = 2000
+
+INITIAL_SEARCH_DECOMPOSITION_ENABLED = True
+ALLOW_REFINEMENT = True
+
 AGENT_DEFAULT_RETRIEVAL_HITS = 15
 AGENT_DEFAULT_RERANKING_HITS = 10
 AGENT_DEFAULT_SUB_QUESTION_MAX_CONTEXT_HITS = 8
@@ -13,9 +31,21 @@ AGENT_DEFAULT_MIN_ORIG_QUESTION_DOCS = 3
 AGENT_DEFAULT_MAX_ANSWER_CONTEXT_DOCS = 10
 AGENT_DEFAULT_MAX_STATIC_HISTORY_WORD_LENGTH = 2000
 
-#####
-# Agent Configs
-#####
+AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_GENERAL_GENERATION = 30  # in seconds
+
+AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_HISTORY_SUMMARY_GENERATION = 10  # in seconds
+AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_ENTITY_TERM_EXTRACTION = 25  # in seconds
+AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_QUERY_REWRITING_GENERATION = 4  # in seconds
+AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_DOCUMENT_VERIFICATION = 1  # in seconds
+AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_SUBQUESTION_GENERATION = 3  # in seconds
+AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_SUBANSWER_GENERATION = 12  # in seconds
+AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_SUBANSWER_CHECK = 8  # in seconds
+AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_INITIAL_ANSWER_GENERATION = 25  # in seconds
+
+AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_REFINED_SUBQUESTION_GENERATION = 6  # in seconds
+AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_REFINED_ANSWER_GENERATION = 25  # in seconds
+AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_REFINED_ANSWER_VALIDATION = 8  # in seconds
+AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_COMPARE_ANSWERS = 8  # in seconds
 
 
 AGENT_RETRIEVAL_STATS = (
@@ -77,4 +107,151 @@ AGENT_MAX_STATIC_HISTORY_WORD_LENGTH = int(
     or AGENT_DEFAULT_MAX_STATIC_HISTORY_WORD_LENGTH
 )  # 2000
 
+AGENT_MAX_STREAMED_DOCS_FOR_INITIAL_ANSWER = int(
+    os.environ.get("AGENT_MAX_STREAMED_DOCS_FOR_INITIAL_ANSWER")
+    or AGENT_DEFAULT_MAX_STREAMED_DOCS_FOR_INITIAL_ANSWER
+)  # 25
+
+AGENT_MAX_STREAMED_DOCS_FOR_REFINED_ANSWER = int(
+    os.environ.get("AGENT_MAX_STREAMED_DOCS_FOR_REFINED_ANSWER")
+    or AGENT_DEFAULT_MAX_STREAMED_DOCS_FOR_REFINED_ANSWER
+)  # 35
+
+
+AGENT_RETRIEVAL_STATS = (
+    not os.environ.get("AGENT_RETRIEVAL_STATS") == "False"
+) or True  # default True
+
+
+AGENT_MAX_QUERY_RETRIEVAL_RESULTS = int(
+    os.environ.get("AGENT_MAX_QUERY_RETRIEVAL_RESULTS") or AGENT_DEFAULT_RETRIEVAL_HITS
+)  # 15
+
+AGENT_MAX_QUERY_RETRIEVAL_RESULTS = int(
+    os.environ.get("AGENT_MAX_QUERY_RETRIEVAL_RESULTS") or AGENT_DEFAULT_RETRIEVAL_HITS
+)  # 15
+
+# Reranking agent configs
+# Reranking stats - no influence on flow outside of stats collection
+AGENT_RERANKING_STATS = (
+    not os.environ.get("AGENT_RERANKING_STATS") == "True"
+) or False  # default False
+
+AGENT_MAX_QUERY_RETRIEVAL_RESULTS = int(
+    os.environ.get("AGENT_MAX_QUERY_RETRIEVAL_RESULTS") or AGENT_DEFAULT_RETRIEVAL_HITS
+)  # 15
+
+AGENT_RERANKING_MAX_QUERY_RETRIEVAL_RESULTS = int(
+    os.environ.get("AGENT_RERANKING_MAX_QUERY_RETRIEVAL_RESULTS")
+    or AGENT_DEFAULT_RERANKING_HITS
+)  # 10
+
+AGENT_NUM_DOCS_FOR_DECOMPOSITION = int(
+    os.environ.get("AGENT_NUM_DOCS_FOR_DECOMPOSITION")
+    or AGENT_DEFAULT_NUM_DOCS_FOR_INITIAL_DECOMPOSITION
+)  # 3
+
+AGENT_NUM_DOCS_FOR_REFINED_DECOMPOSITION = int(
+    os.environ.get("AGENT_NUM_DOCS_FOR_REFINED_DECOMPOSITION")
+    or AGENT_DEFAULT_NUM_DOCS_FOR_REFINED_DECOMPOSITION
+)  # 5
+
+AGENT_EXPLORATORY_SEARCH_RESULTS = int(
+    os.environ.get("AGENT_EXPLORATORY_SEARCH_RESULTS")
+    or AGENT_DEFAULT_EXPLORATORY_SEARCH_RESULTS
+)  # 5
+
+AGENT_MIN_ORIG_QUESTION_DOCS = int(
+    os.environ.get("AGENT_MIN_ORIG_QUESTION_DOCS")
+    or AGENT_DEFAULT_MIN_ORIG_QUESTION_DOCS
+)  # 3
+
+AGENT_MAX_ANSWER_CONTEXT_DOCS = int(
+    os.environ.get("AGENT_MAX_ANSWER_CONTEXT_DOCS")
+    or AGENT_DEFAULT_SUB_QUESTION_MAX_CONTEXT_HITS
+)  # 8
+
+
+AGENT_MAX_STATIC_HISTORY_WORD_LENGTH = int(
+    os.environ.get("AGENT_MAX_STATIC_HISTORY_WORD_LENGTH")
+    or AGENT_DEFAULT_MAX_STATIC_HISTORY_WORD_LENGTH
+)  # 2000
+
+
+AGENT_TIMEOUT_OVERRIDE_LLM_ENTITY_TERM_EXTRACTION = int(
+    os.environ.get("AGENT_TIMEOUT_OVERRIDE_LLM_ENTITY_TERM_EXTRACTION")
+    or AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_ENTITY_TERM_EXTRACTION
+)  # 25
+
+
+AGENT_TIMEOUT_OVERRIDE_LLM_DOCUMENT_VERIFICATION = int(
+    os.environ.get("AGENT_TIMEOUT_OVERRIDE_LLM_DOCUMENT_VERIFICATION")
+    or AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_DOCUMENT_VERIFICATION
+)  # 3
+
+AGENT_TIMEOUT_OVERRIDE_LLM_GENERAL_GENERATION = int(
+    os.environ.get("AGENT_TIMEOUT_OVERRIDE_LLM_GENERAL_GENERATION")
+    or AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_GENERAL_GENERATION
+)  # 30
+
+
+AGENT_TIMEOUT_OVERRIDE_LLM_SUBQUESTION_GENERATION = int(
+    os.environ.get("AGENT_TIMEOUT_OVERRIDE_LLM_SUBQUESTION_GENERATION")
+    or AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_SUBQUESTION_GENERATION
+)  # 8
+
+
+AGENT_TIMEOUT_OVERRIDE_LLM_SUBANSWER_GENERATION = int(
+    os.environ.get("AGENT_TIMEOUT_OVERRIDE_LLM_SUBANSWER_GENERATION")
+    or AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_SUBANSWER_GENERATION
+)  # 12
+
+
+AGENT_TIMEOUT_OVERRIDE_LLM_INITIAL_ANSWER_GENERATION = int(
+    os.environ.get("AGENT_TIMEOUT_OVERRIDE_LLM_INITIAL_ANSWER_GENERATION")
+    or AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_INITIAL_ANSWER_GENERATION
+)  # 25
+
+
+AGENT_TIMEOUT_OVERRIDE_LLM_REFINED_ANSWER_GENERATION = int(
+    os.environ.get("AGENT_TIMEOUT_OVERRIDE_LLM_REFINED_ANSWER_GENERATION")
+    or AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_REFINED_ANSWER_GENERATION
+)  # 25
+
+
+AGENT_TIMEOUT_OVERRIDE_LLM_SUBANSWER_CHECK = int(
+    os.environ.get("AGENT_TIMEOUT_OVERRIDE_LLM_SUBANSWER_CHECK")
+    or AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_SUBANSWER_CHECK
+)  # 8
+
+
+AGENT_TIMEOUT_OVERRIDE_LLM_REFINED_SUBQUESTION_GENERATION = int(
+    os.environ.get("AGENT_TIMEOUT_OVERRIDE_LLM_REFINED_SUBQUESTION_GENERATION")
+    or AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_REFINED_SUBQUESTION_GENERATION
+)  # 6
+
+
+AGENT_TIMEOUT_OVERRIDE_LLM_QUERY_REWRITING_GENERATION = int(
+    os.environ.get("AGENT_TIMEOUT_OVERRIDE_LLM_QUERY_REWRITING_GENERATION")
+    or AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_QUERY_REWRITING_GENERATION
+)  # 1
+
+
+AGENT_TIMEOUT_OVERRIDE_LLM_HISTORY_SUMMARY_GENERATION = int(
+    os.environ.get("AGENT_TIMEOUT_OVERRIDE_LLM_HISTORY_SUMMARY_GENERATION")
+    or AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_HISTORY_SUMMARY_GENERATION
+)  # 4
+
+
+AGENT_TIMEOUT_OVERRIDE_LLM_COMPARE_ANSWERS = int(
+    os.environ.get("AGENT_TIMEOUT_OVERRIDE_LLM_COMPARE_ANSWERS")
+    or AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_COMPARE_ANSWERS
+)  # 8
+
+
+AGENT_TIMEOUT_OVERRIDE_LLM_REFINED_ANSWER_VALIDATION = int(
+    os.environ.get("AGENT_TIMEOUT_OVERRIDE_LLM_REFINED_ANSWER_VALIDATION")
+    or AGENT_DEFAULT_TIMEOUT_OVERRIDE_LLM_REFINED_ANSWER_VALIDATION
+)  # 8
+
 GRAPH_VERSION_NAME: str = "a"

backend/onyx/configs/constants.py

@@ -107,9 +107,9 @@ CELERY_TASK_WAIT_FOR_FENCE_TIMEOUT = 5 * 60  # 5 min
 
 # needs to be long enough to cover the maximum time it takes to download an object
 # if we can get callbacks as object bytes download, we could lower this a lot.
-CELERY_PRUNING_LOCK_TIMEOUT = 300  # 5 min
+CELERY_PRUNING_LOCK_TIMEOUT = 3600  # 1 hour (in seconds)
 
-CELERY_PERMISSIONS_SYNC_LOCK_TIMEOUT = 300  # 5 min
+CELERY_PERMISSIONS_SYNC_LOCK_TIMEOUT = 3600  # 1 hour (in seconds)
 
 CELERY_EXTERNAL_GROUP_SYNC_LOCK_TIMEOUT = 300  # 5 min
 
@@ -125,6 +125,7 @@ class DocumentSource(str, Enum):
     GMAIL = "gmail"
     REQUESTTRACKER = "requesttracker"
     GITHUB = "github"
+    GITBOOK = "gitbook"
     GITLAB = "gitlab"
     GURU = "guru"
     BOOKSTACK = "bookstack"
@@ -298,7 +299,6 @@ class OnyxRedisLocks:
     CHECK_CONNECTOR_EXTERNAL_GROUP_SYNC_BEAT_LOCK = (
         "da_lock:check_connector_external_group_sync_beat"
     )
-    MONITOR_VESPA_SYNC_BEAT_LOCK = "da_lock:monitor_vespa_sync_beat"
     MONITOR_BACKGROUND_PROCESSES_LOCK = "da_lock:monitor_background_processes"
 
     CONNECTOR_DOC_PERMISSIONS_SYNC_LOCK_PREFIX = (
@@ -324,6 +324,7 @@ class OnyxRedisSignals:
     BLOCK_VALIDATE_PERMISSION_SYNC_FENCES = (
         "signal:block_validate_permission_sync_fences"
     )
+    BLOCK_PRUNING = "signal:block_pruning"
     BLOCK_VALIDATE_PRUNING_FENCES = "signal:block_validate_pruning_fences"
     BLOCK_BUILD_FENCE_LOOKUP_TABLE = "signal:block_build_fence_lookup_table"
 
@@ -346,12 +347,18 @@ ONYX_CLOUD_CELERY_TASK_PREFIX = "cloud"
 # the tenant id we use for system level redis operations
 ONYX_CLOUD_TENANT_ID = "cloud"
 
+# the redis namespace for runtime variables
+ONYX_CLOUD_REDIS_RUNTIME = "runtime"
+
 
 class OnyxCeleryTask:
     DEFAULT = "celery"
 
     CLOUD_BEAT_TASK_GENERATOR = f"{ONYX_CLOUD_CELERY_TASK_PREFIX}_generate_beat_tasks"
-    CLOUD_CHECK_ALEMBIC = f"{ONYX_CLOUD_CELERY_TASK_PREFIX}_check_alembic"
+    CLOUD_MONITOR_ALEMBIC = f"{ONYX_CLOUD_CELERY_TASK_PREFIX}_monitor_alembic"
+    CLOUD_MONITOR_CELERY_QUEUES = (
+        f"{ONYX_CLOUD_CELERY_TASK_PREFIX}_monitor_celery_queues"
+    )
 
     CHECK_FOR_CONNECTOR_DELETION = "check_for_connector_deletion_task"
     CHECK_FOR_VESPA_SYNC_TASK = "check_for_vespa_sync_task"
@@ -361,8 +368,8 @@ class OnyxCeleryTask:
     CHECK_FOR_EXTERNAL_GROUP_SYNC = "check_for_external_group_sync"
     CHECK_FOR_LLM_MODEL_UPDATE = "check_for_llm_model_update"
 
-    MONITOR_VESPA_SYNC = "monitor_vespa_sync"
     MONITOR_BACKGROUND_PROCESSES = "monitor_background_processes"
+    MONITOR_CELERY_QUEUES = "monitor_celery_queues"
 
     KOMBU_MESSAGE_CLEANUP_TASK = "kombu_message_cleanup_task"
     CONNECTOR_PERMISSION_SYNC_GENERATOR_TASK = (

backend/onyx/connectors/airtable/airtable_connector.py

@@ -65,10 +65,25 @@ class AirtableConnector(LoadConnector):
         base_id: str,
         table_name_or_id: str,
         treat_all_non_attachment_fields_as_metadata: bool = False,
+        view_id: str | None = None,
+        share_id: str | None = None,
         batch_size: int = INDEX_BATCH_SIZE,
     ) -> None:
+        """Initialize an AirtableConnector.
+
+        Args:
+            base_id: The ID of the Airtable base to connect to
+            table_name_or_id: The name or ID of the table to index
+            treat_all_non_attachment_fields_as_metadata: If True, all fields except attachments will be treated as metadata.
+                If False, only fields with types in DEFAULT_METADATA_FIELD_TYPES will be treated as metadata.
+            view_id: Optional ID of a specific view to use
+            share_id: Optional ID of a "share" to use for generating record URLs (https://airtable.com/developers/web/api/list-shares)
+            batch_size: Number of records to process in each batch
+        """
         self.base_id = base_id
         self.table_name_or_id = table_name_or_id
+        self.view_id = view_id
+        self.share_id = share_id
         self.batch_size = batch_size
         self._airtable_client: AirtableApi | None = None
         self.treat_all_non_attachment_fields_as_metadata = (
@@ -85,6 +100,39 @@ class AirtableConnector(LoadConnector):
             raise AirtableClientNotSetUpError()
         return self._airtable_client
 
+    @classmethod
+    def _get_record_url(
+        cls,
+        base_id: str,
+        table_id: str,
+        record_id: str,
+        share_id: str | None,
+        view_id: str | None,
+        field_id: str | None = None,
+        attachment_id: str | None = None,
+    ) -> str:
+        """Constructs the URL for a record, optionally including field and attachment IDs
+
+        Full possible structure is:
+
+        https://airtable.com/BASE_ID/SHARE_ID/TABLE_ID/VIEW_ID/RECORD_ID/FIELD_ID/ATTACHMENT_ID
+        """
+        # If we have a shared link, use that view for better UX
+        if share_id:
+            base_url = f"https://airtable.com/{base_id}/{share_id}/{table_id}"
+        else:
+            base_url = f"https://airtable.com/{base_id}/{table_id}"
+
+        if view_id:
+            base_url = f"{base_url}/{view_id}"
+
+        base_url = f"{base_url}/{record_id}"
+
+        if field_id and attachment_id:
+            return f"{base_url}/{field_id}/{attachment_id}?blocks=hide"
+
+        return base_url
+
     def _extract_field_values(
         self,
         field_id: str,
@@ -110,8 +158,10 @@ class AirtableConnector(LoadConnector):
         if field_type == "multipleRecordLinks":
             return []
 
-        # default link to use for non-attachment fields
-        default_link = f"https://airtable.com/{base_id}/{table_id}/{record_id}"
+        # Get the base URL for this record
+        default_link = self._get_record_url(
+            base_id, table_id, record_id, self.share_id, self.view_id or view_id
+        )
 
         if field_type == "multipleAttachments":
             attachment_texts: list[tuple[str, str]] = []
@@ -165,17 +215,16 @@ class AirtableConnector(LoadConnector):
                             extension=file_ext,
                         )
                         if attachment_text:
-                            # slightly nicer loading experience if we can specify the view ID
-                            if view_id:
-                                attachment_link = (
-                                    f"https://airtable.com/{base_id}/{table_id}/{view_id}/{record_id}"
-                                    f"/{field_id}/{attachment_id}?blocks=hide"
-                                )
-                            else:
-                                attachment_link = (
-                                    f"https://airtable.com/{base_id}/{table_id}/{record_id}"
-                                    f"/{field_id}/{attachment_id}?blocks=hide"
-                                )
+                            # Use the helper method to construct attachment URLs
+                            attachment_link = self._get_record_url(
+                                base_id,
+                                table_id,
+                                record_id,
+                                self.share_id,
+                                self.view_id or view_id,
+                                field_id,
+                                attachment_id,
+                            )
                             attachment_texts.append(
                                 (f"{filename}:\n{attachment_text}", attachment_link)
                             )

backend/onyx/connectors/factory.py

@@ -20,6 +20,7 @@ from onyx.connectors.egnyte.connector import EgnyteConnector
 from onyx.connectors.file.connector import LocalFileConnector
 from onyx.connectors.fireflies.connector import FirefliesConnector
 from onyx.connectors.freshdesk.connector import FreshdeskConnector
+from onyx.connectors.gitbook.connector import GitbookConnector
 from onyx.connectors.github.connector import GithubConnector
 from onyx.connectors.gitlab.connector import GitlabConnector
 from onyx.connectors.gmail.connector import GmailConnector
@@ -71,6 +72,7 @@ def identify_connector_class(
         DocumentSource.GITHUB: GithubConnector,
         DocumentSource.GMAIL: GmailConnector,
         DocumentSource.GITLAB: GitlabConnector,
+        DocumentSource.GITBOOK: GitbookConnector,
         DocumentSource.GOOGLE_DRIVE: GoogleDriveConnector,
         DocumentSource.BOOKSTACK: BookstackConnector,
         DocumentSource.CONFLUENCE: ConfluenceConnector,

backend/onyx/connectors/gitbook/connector.py

@@ -0,0 +1,279 @@
+from datetime import datetime
+from datetime import timezone
+from typing import Any
+from urllib.parse import urljoin
+
+import requests
+
+from onyx.configs.app_configs import INDEX_BATCH_SIZE
+from onyx.configs.constants import DocumentSource
+from onyx.connectors.interfaces import GenerateDocumentsOutput
+from onyx.connectors.interfaces import LoadConnector
+from onyx.connectors.interfaces import PollConnector
+from onyx.connectors.interfaces import SecondsSinceUnixEpoch
+from onyx.connectors.models import ConnectorMissingCredentialError
+from onyx.connectors.models import Document
+from onyx.connectors.models import Section
+from onyx.utils.logger import setup_logger
+
+
+logger = setup_logger()
+
+GITBOOK_API_BASE = "https://api.gitbook.com/v1/"
+
+
+class GitbookApiClient:
+    def __init__(self, access_token: str) -> None:
+        self.access_token = access_token
+
+    def get(self, endpoint: str, params: dict[str, Any] | None = None) -> Any:
+        headers = {
+            "Authorization": f"Bearer {self.access_token}",
+            "Content-Type": "application/json",
+        }
+
+        url = urljoin(GITBOOK_API_BASE, endpoint.lstrip("/"))
+        response = requests.get(url, headers=headers, params=params)
+        response.raise_for_status()
+        return response.json()
+
+    def get_page_content(self, space_id: str, page_id: str) -> dict[str, Any]:
+        return self.get(f"/spaces/{space_id}/content/page/{page_id}")
+
+
+def _extract_text_from_document(document: dict[str, Any]) -> str:
+    """Extract text content from GitBook document structure by parsing the document nodes
+    into markdown format."""
+
+    def parse_leaf(leaf: dict[str, Any]) -> str:
+        text = leaf.get("text", "")
+        leaf.get("marks", [])
+        return text
+
+    def parse_text_node(node: dict[str, Any]) -> str:
+        text = ""
+        for leaf in node.get("leaves", []):
+            text += parse_leaf(leaf)
+        return text
+
+    def parse_block_node(node: dict[str, Any]) -> str:
+        block_type = node.get("type", "")
+        result = ""
+
+        if block_type == "heading-1":
+            text = "".join(parse_text_node(n) for n in node.get("nodes", []))
+            result = f"# {text}\n\n"
+
+        elif block_type == "heading-2":
+            text = "".join(parse_text_node(n) for n in node.get("nodes", []))
+            result = f"## {text}\n\n"
+
+        elif block_type == "heading-3":
+            text = "".join(parse_text_node(n) for n in node.get("nodes", []))
+            result = f"### {text}\n\n"
+
+        elif block_type == "heading-4":
+            text = "".join(parse_text_node(n) for n in node.get("nodes", []))
+            result = f"#### {text}\n\n"
+
+        elif block_type == "heading-5":
+            text = "".join(parse_text_node(n) for n in node.get("nodes", []))
+            result = f"##### {text}\n\n"
+
+        elif block_type == "heading-6":
+            text = "".join(parse_text_node(n) for n in node.get("nodes", []))
+            result = f"###### {text}\n\n"
+
+        elif block_type == "list-unordered":
+            for list_item in node.get("nodes", []):
+                paragraph = list_item.get("nodes", [])[0]
+                text = "".join(parse_text_node(n) for n in paragraph.get("nodes", []))
+                result += f"* {text}\n"
+            result += "\n"
+
+        elif block_type == "paragraph":
+            text = "".join(parse_text_node(n) for n in node.get("nodes", []))
+            result = f"{text}\n\n"
+
+        elif block_type == "list-tasks":
+            for task_item in node.get("nodes", []):
+                checked = task_item.get("data", {}).get("checked", False)
+                paragraph = task_item.get("nodes", [])[0]
+                text = "".join(parse_text_node(n) for n in paragraph.get("nodes", []))
+                checkbox = "[x]" if checked else "[ ]"
+                result += f"- {checkbox} {text}\n"
+            result += "\n"
+
+        elif block_type == "code":
+            for code_line in node.get("nodes", []):
+                if code_line.get("type") == "code-line":
+                    text = "".join(
+                        parse_text_node(n) for n in code_line.get("nodes", [])
+                    )
+                    result += f"{text}\n"
+            result += "\n"
+
+        elif block_type == "blockquote":
+            for quote_node in node.get("nodes", []):
+                if quote_node.get("type") == "paragraph":
+                    text = "".join(
+                        parse_text_node(n) for n in quote_node.get("nodes", [])
+                    )
+                    result += f"> {text}\n"
+            result += "\n"
+
+        elif block_type == "table":
+            records = node.get("data", {}).get("records", {})
+            definition = node.get("data", {}).get("definition", {})
+            view = node.get("data", {}).get("view", {})
+
+            columns = view.get("columns", [])
+
+            header_cells = []
+            for col_id in columns:
+                col_def = definition.get(col_id, {})
+                header_cells.append(col_def.get("title", ""))
+
+            result = "| " + " | ".join(header_cells) + " |\n"
+            result += "|" + "---|" * len(header_cells) + "\n"
+
+            sorted_records = sorted(
+                records.items(), key=lambda x: x[1].get("orderIndex", "")
+            )
+
+            for record_id, record_data in sorted_records:
+                values = record_data.get("values", {})
+                row_cells = []
+                for col_id in columns:
+                    fragment_id = values.get(col_id, "")
+                    fragment_text = ""
+                    for fragment in node.get("fragments", []):
+                        if fragment.get("fragment") == fragment_id:
+                            for frag_node in fragment.get("nodes", []):
+                                if frag_node.get("type") == "paragraph":
+                                    fragment_text = "".join(
+                                        parse_text_node(n)
+                                        for n in frag_node.get("nodes", [])
+                                    )
+                                    break
+                    row_cells.append(fragment_text)
+                result += "| " + " | ".join(row_cells) + " |\n"
+
+            result += "\n"
+        return result
+
+    if not document or "document" not in document:
+        return ""
+
+    markdown = ""
+    nodes = document["document"].get("nodes", [])
+
+    for node in nodes:
+        markdown += parse_block_node(node)
+
+    return markdown
+
+
+def _convert_page_to_document(
+    client: GitbookApiClient, space_id: str, page: dict[str, Any]
+) -> Document:
+    page_id = page["id"]
+    page_content = client.get_page_content(space_id, page_id)
+
+    return Document(
+        id=f"gitbook-{space_id}-{page_id}",
+        sections=[
+            Section(
+                link=page.get("urls", {}).get("app", ""),
+                text=_extract_text_from_document(page_content),
+            )
+        ],
+        source=DocumentSource.GITBOOK,
+        semantic_identifier=page.get("title", ""),
+        doc_updated_at=datetime.fromisoformat(page["updatedAt"]).replace(
+            tzinfo=timezone.utc
+        ),
+        metadata={
+            "path": page.get("path", ""),
+            "type": page.get("type", ""),
+            "kind": page.get("kind", ""),
+        },
+    )
+
+
+class GitbookConnector(LoadConnector, PollConnector):
+    def __init__(
+        self,
+        space_id: str,
+        batch_size: int = INDEX_BATCH_SIZE,
+    ) -> None:
+        self.space_id = space_id
+        self.batch_size = batch_size
+        self.access_token: str | None = None
+        self.client: GitbookApiClient | None = None
+
+    def load_credentials(self, credentials: dict[str, Any]) -> None:
+        access_token = credentials.get("gitbook_api_key")
+        if not access_token:
+            raise ConnectorMissingCredentialError("GitBook access token")
+        self.access_token = access_token
+        self.client = GitbookApiClient(access_token)
+
+    def _fetch_all_pages(
+        self,
+        start: datetime | None = None,
+        end: datetime | None = None,
+    ) -> GenerateDocumentsOutput:
+        if not self.client:
+            raise ConnectorMissingCredentialError("GitBook")
+
+        try:
+            content = self.client.get(f"/spaces/{self.space_id}/content")
+            pages = content.get("pages", [])
+
+            current_batch: list[Document] = []
+            for page in pages:
+                updated_at = datetime.fromisoformat(page["updatedAt"])
+
+                if start and updated_at < start:
+                    if current_batch:
+                        yield current_batch
+                    return
+                if end and updated_at > end:
+                    continue
+
+                current_batch.append(
+                    _convert_page_to_document(self.client, self.space_id, page)
+                )
+
+                if len(current_batch) >= self.batch_size:
+                    yield current_batch
+                    current_batch = []
+
+            if current_batch:
+                yield current_batch
+
+        except requests.RequestException as e:
+            logger.error(f"Error fetching GitBook content: {str(e)}")
+            raise
+
+    def load_from_state(self) -> GenerateDocumentsOutput:
+        return self._fetch_all_pages()
+
+    def poll_source(
+        self, start: SecondsSinceUnixEpoch, end: SecondsSinceUnixEpoch
+    ) -> GenerateDocumentsOutput:
+        start_datetime = datetime.fromtimestamp(start, tz=timezone.utc)
+        end_datetime = datetime.fromtimestamp(end, tz=timezone.utc)
+        return self._fetch_all_pages(start_datetime, end_datetime)
+
+
+if __name__ == "__main__":
+    import os
+
+    connector = GitbookConnector(
+        space_id=os.environ["GITBOOK_SPACE_ID"],
+    )
+    connector.load_credentials({"gitbook_api_key": os.environ["GITBOOK_API_KEY"]})
+    document_batches = connector.load_from_state()
+    print(next(document_batches))

backend/onyx/connectors/google_drive/connector.py

@@ -302,7 +302,7 @@ class GoogleDriveConnector(LoadConnector, PollConnector, SlimConnector):
             if e.status_code == 401:
                 # fail gracefully, let the other impersonations continue
                 # one user without access shouldn't block the entire connector
-                logger.exception(
+                logger.warning(
                     f"User '{user_email}' does not have access to the drive APIs."
                 )
                 return

backend/onyx/connectors/onyx_jira/connector.py

@@ -145,7 +145,8 @@ def fetch_jira_issues_batch(
             id=page_url,
             sections=[Section(link=page_url, text=ticket_content)],
             source=DocumentSource.JIRA,
-            semantic_identifier=issue.fields.summary,
+            semantic_identifier=f"{issue.key}: {issue.fields.summary}",
+            title=f"{issue.key} {issue.fields.summary}",
             doc_updated_at=time_str_to_utc(issue.fields.updated),
             primary_owners=list(people) or None,
             # TODO add secondary_owners (commenters) if needed

backend/onyx/connectors/slack/utils.py

@@ -39,19 +39,6 @@ def get_message_link(
     return permalink
 
 
-def _make_slack_api_call_logged(
-    call: Callable[..., SlackResponse],
-) -> Callable[..., SlackResponse]:
-    @wraps(call)
-    def logged_call(**kwargs: Any) -> SlackResponse:
-        logger.debug(f"Making call to Slack API '{call.__name__}' with args '{kwargs}'")
-        result = call(**kwargs)
-        logger.debug(f"Call to Slack API '{call.__name__}' returned '{result}'")
-        return result
-
-    return logged_call
-
-
 def _make_slack_api_call_paginated(
     call: Callable[..., SlackResponse],
 ) -> Callable[..., Generator[dict[str, Any], None, None]]:
@@ -127,18 +114,14 @@ def make_slack_api_rate_limited(
 def make_slack_api_call_w_retries(
     call: Callable[..., SlackResponse], **kwargs: Any
 ) -> SlackResponse:
-    return basic_retry_wrapper(
-        make_slack_api_rate_limited(_make_slack_api_call_logged(call))
-    )(**kwargs)
+    return basic_retry_wrapper(make_slack_api_rate_limited(call))(**kwargs)
 
 
 def make_paginated_slack_api_call_w_retries(
     call: Callable[..., SlackResponse], **kwargs: Any
 ) -> Generator[dict[str, Any], None, None]:
     return _make_slack_api_call_paginated(
-        basic_retry_wrapper(
-            make_slack_api_rate_limited(_make_slack_api_call_logged(call))
-        )
+        basic_retry_wrapper(make_slack_api_rate_limited(call))
     )(**kwargs)
 
 

backend/onyx/context/search/pipeline.py

@@ -51,6 +51,7 @@ class SearchPipeline:
         user: User | None,
         llm: LLM,
         fast_llm: LLM,
+        skip_query_analysis: bool,
         db_session: Session,
         bypass_acl: bool = False,  # NOTE: VERY DANGEROUS, USE WITH CAUTION
         retrieval_metrics_callback: (
@@ -61,10 +62,13 @@ class SearchPipeline:
         rerank_metrics_callback: Callable[[RerankMetricsContainer], None] | None = None,
         prompt_config: PromptConfig | None = None,
     ):
+        # NOTE: The Search Request contains a lot of fields that are overrides, many of them can be None
+        # and typically are None. The preprocessing will fetch default values to replace these empty overrides.
         self.search_request = search_request
         self.user = user
         self.llm = llm
         self.fast_llm = fast_llm
+        self.skip_query_analysis = skip_query_analysis
         self.db_session = db_session
         self.bypass_acl = bypass_acl
         self.retrieval_metrics_callback = retrieval_metrics_callback
@@ -106,6 +110,7 @@ class SearchPipeline:
             search_request=self.search_request,
             user=self.user,
             llm=self.llm,
+            skip_query_analysis=self.skip_query_analysis,
             db_session=self.db_session,
             bypass_acl=self.bypass_acl,
         )
@@ -160,6 +165,12 @@ class SearchPipeline:
         that have a corresponding chunk.
 
         This step should be fast for any document index implementation.
+
+        Current implementation timing is approximately broken down in timing as:
+        - 200 ms to get the embedding of the query
+        - 15 ms to get chunks from the document index
+        - possibly more to get additional surrounding chunks
+        - possibly more for query expansion (multilingual)
         """
         if self._retrieved_sections is not None:
             return self._retrieved_sections

backend/onyx/context/search/postprocessing/postprocessing.py

@@ -15,6 +15,7 @@ from onyx.context.search.models import InferenceChunk
 from onyx.context.search.models import InferenceChunkUncleaned
 from onyx.context.search.models import InferenceSection
 from onyx.context.search.models import MAX_METRICS_CONTENT
+from onyx.context.search.models import RerankingDetails
 from onyx.context.search.models import RerankMetricsContainer
 from onyx.context.search.models import SearchQuery
 from onyx.document_index.document_index_utils import (
@@ -77,7 +78,8 @@ def cleanup_chunks(chunks: list[InferenceChunkUncleaned]) -> list[InferenceChunk
 
 @log_function_time(print_only=True)
 def semantic_reranking(
-    query: SearchQuery,
+    query_str: str,
+    rerank_settings: RerankingDetails,
     chunks: list[InferenceChunk],
     model_min: int = CROSS_ENCODER_RANGE_MIN,
     model_max: int = CROSS_ENCODER_RANGE_MAX,
@@ -88,11 +90,9 @@ def semantic_reranking(
 
     Note: this updates the chunks in place, it updates the chunk scores which came from retrieval
     """
-    rerank_settings = query.rerank_settings
-
-    if not rerank_settings or not rerank_settings.rerank_model_name:
-        # Should never reach this part of the flow without reranking settings
-        raise RuntimeError("Reranking flow should not be running")
+    assert (
+        rerank_settings.rerank_model_name
+    ), "Reranking flow cannot run without a specific model"
 
     chunks_to_rerank = chunks[: rerank_settings.num_rerank]
 
@@ -107,7 +107,7 @@ def semantic_reranking(
         f"{chunk.semantic_identifier or chunk.title or ''}\n{chunk.content}"
         for chunk in chunks_to_rerank
     ]
-    sim_scores_floats = cross_encoder.predict(query=query.query, passages=passages)
+    sim_scores_floats = cross_encoder.predict(query=query_str, passages=passages)
 
     # Old logic to handle multiple cross-encoders preserved but not used
     sim_scores = [numpy.array(sim_scores_floats)]
@@ -165,8 +165,20 @@ def semantic_reranking(
     return list(ranked_chunks), list(ranked_indices)
 
 
+def should_rerank(rerank_settings: RerankingDetails | None) -> bool:
+    """Based on the RerankingDetails model, only run rerank if the following conditions are met:
+    - rerank_model_name is not None
+    - num_rerank is greater than 0
+    """
+    if not rerank_settings:
+        return False
+
+    return bool(rerank_settings.rerank_model_name and rerank_settings.num_rerank > 0)
+
+
 def rerank_sections(
-    query: SearchQuery,
+    query_str: str,
+    rerank_settings: RerankingDetails,
     sections_to_rerank: list[InferenceSection],
     rerank_metrics_callback: Callable[[RerankMetricsContainer], None] | None = None,
 ) -> list[InferenceSection]:
@@ -181,16 +193,13 @@ def rerank_sections(
     """
     chunks_to_rerank = [section.center_chunk for section in sections_to_rerank]
 
-    if not query.rerank_settings:
-        # Should never reach this part of the flow without reranking settings
-        raise RuntimeError("Reranking settings not found")
-
     ranked_chunks, _ = semantic_reranking(
-        query=query,
+        query_str=query_str,
+        rerank_settings=rerank_settings,
         chunks=chunks_to_rerank,
         rerank_metrics_callback=rerank_metrics_callback,
     )
-    lower_chunks = chunks_to_rerank[query.rerank_settings.num_rerank :]
+    lower_chunks = chunks_to_rerank[rerank_settings.num_rerank :]
 
     # Scores from rerank cannot be meaningfully combined with scores without rerank
     # However the ordering is still important
@@ -260,16 +269,13 @@ def search_postprocessing(
 
     rerank_task_id = None
     sections_yielded = False
-    if (
-        search_query.rerank_settings
-        and search_query.rerank_settings.rerank_model_name
-        and search_query.rerank_settings.num_rerank > 0
-    ):
+    if should_rerank(search_query.rerank_settings):
         post_processing_tasks.append(
             FunctionCall(
                 rerank_sections,
                 (
-                    search_query,
+                    search_query.query,
+                    search_query.rerank_settings,  # Cannot be None here
                     retrieved_sections,
                     rerank_metrics_callback,
                 ),

backend/onyx/context/search/preprocessing/preprocessing.py

@@ -50,11 +50,11 @@ def retrieval_preprocessing(
     search_request: SearchRequest,
     user: User | None,
     llm: LLM,
+    skip_query_analysis: bool,
     db_session: Session,
-    bypass_acl: bool = False,
-    skip_query_analysis: bool = False,
-    base_recency_decay: float = BASE_RECENCY_DECAY,
     favor_recent_decay_multiplier: float = FAVOR_RECENT_DECAY_MULTIPLIER,
+    base_recency_decay: float = BASE_RECENCY_DECAY,
+    bypass_acl: bool = False,
 ) -> SearchQuery:
     """Logic is as follows:
     Any global disables apply first
@@ -146,7 +146,7 @@ def retrieval_preprocessing(
     is_keyword, extracted_keywords = (
         parallel_results[run_query_analysis.result_id]
         if run_query_analysis
-        else (None, None)
+        else (False, None)
     )
 
     all_query_terms = query.split()

backend/onyx/db/background_error.py

@@ -0,0 +1,10 @@
+from sqlalchemy.orm import Session
+
+from onyx.db.models import BackgroundError
+
+
+def create_background_error(
+    db_session: Session, message: str, cc_pair_id: int | None
+) -> None:
+    db_session.add(BackgroundError(message=message, cc_pair_id=cc_pair_id))
+    db_session.commit()

backend/onyx/db/chat.py

@@ -350,13 +350,14 @@ def delete_chat_session(
     user_id: UUID | None,
     chat_session_id: UUID,
     db_session: Session,
+    include_deleted: bool = False,
     hard_delete: bool = HARD_DELETE_CHATS,
 ) -> None:
     chat_session = get_chat_session_by_id(
         chat_session_id=chat_session_id, user_id=user_id, db_session=db_session
     )
 
-    if chat_session.deleted:
+    if chat_session.deleted and not include_deleted:
         raise ValueError("Cannot delete an already deleted chat session")
 
     if hard_delete:
@@ -380,7 +381,15 @@ def delete_chat_sessions_older_than(days_old: int, db_session: Session) -> None:
     ).fetchall()
 
     for user_id, session_id in old_sessions:
-        delete_chat_session(user_id, session_id, db_session, hard_delete=True)
+        try:
+            delete_chat_session(
+                user_id, session_id, db_session, include_deleted=True, hard_delete=True
+            )
+        except Exception:
+            logger.exception(
+                "delete_chat_session exceptioned. "
+                f"user_id={user_id} session_id={session_id}"
+            )
 
 
 def get_chat_message(
@@ -893,14 +902,18 @@ def translate_db_sub_questions_to_server_objects(
                 question=sub_question.sub_question,
                 answer=sub_question.sub_answer,
                 sub_queries=sub_queries,
-                context_docs=get_retrieval_docs_from_search_docs(verified_docs),
+                context_docs=get_retrieval_docs_from_search_docs(
+                    verified_docs, sort_by_score=False
+                ),
             )
         )
     return sub_questions
 
 
 def get_retrieval_docs_from_search_docs(
-    search_docs: list[SearchDoc], remove_doc_content: bool = False
+    search_docs: list[SearchDoc],
+    remove_doc_content: bool = False,
+    sort_by_score: bool = True,
 ) -> RetrievalDocs:
     top_documents = [
         translate_db_search_doc_to_server_search_doc(
@@ -908,7 +921,8 @@ def get_retrieval_docs_from_search_docs(
         )
         for db_doc in search_docs
     ]
-    top_documents = sorted(top_documents, key=lambda doc: doc.score, reverse=True)  # type: ignore
+    if sort_by_score:
+        top_documents = sorted(top_documents, key=lambda doc: doc.score, reverse=True)  # type: ignore
     return RetrievalDocs(top_documents=top_documents)
 
 
@@ -1018,7 +1032,7 @@ def log_agent_sub_question_results(
         sub_question = sub_question_answer_result.question
         sub_answer = sub_question_answer_result.answer
         sub_document_results = _create_citation_format_list(
-            sub_question_answer_result.verified_reranked_documents
+            sub_question_answer_result.context_documents
         )
 
         sub_question_object = AgentSubQuestion(

backend/onyx/db/models.py

@@ -483,6 +483,10 @@ class ConnectorCredentialPair(Base):
         primaryjoin="foreign(ConnectorCredentialPair.creator_id) == remote(User.id)",
     )
 
+    background_errors: Mapped[list["BackgroundError"]] = relationship(
+        "BackgroundError", back_populates="cc_pair", cascade="all, delete-orphan"
+    )
+
 
 class Document(Base):
     __tablename__ = "document"
@@ -2115,6 +2119,31 @@ class StandardAnswer(Base):
     )
 
 
+class BackgroundError(Base):
+    """Important background errors. Serves to:
+    1. Ensure that important logs are kept around and not lost on rotation/container restarts
+    2. A trail for high-signal events so that the debugger doesn't need to remember/know every
+       possible relevant log line.
+    """
+
+    __tablename__ = "background_error"
+
+    id: Mapped[int] = mapped_column(primary_key=True)
+    message: Mapped[str] = mapped_column(String)
+    time_created: Mapped[datetime.datetime] = mapped_column(
+        DateTime(timezone=True), server_default=func.now()
+    )
+
+    # option to link the error to a specific CC Pair
+    cc_pair_id: Mapped[int | None] = mapped_column(
+        ForeignKey("connector_credential_pair.id", ondelete="CASCADE"), nullable=True
+    )
+
+    cc_pair: Mapped["ConnectorCredentialPair | None"] = relationship(
+        "ConnectorCredentialPair", back_populates="background_errors"
+    )
+
+
 """Tables related to Permission Sync"""
 
 

backend/onyx/db/persona.py

@@ -204,6 +204,14 @@ def create_update_persona(
         if not all_prompt_ids:
             raise ValueError("No prompt IDs provided")
 
+        # Default persona validation
+        if create_persona_request.is_default_persona:
+            if not create_persona_request.is_public:
+                raise ValueError("Cannot make a default persona non public")
+
+            if user and user.role != UserRole.ADMIN:
+                raise ValueError("Only admins can make a default persona")
+
         persona = upsert_persona(
             persona_id=persona_id,
             user=user,
@@ -510,6 +518,7 @@ def upsert_persona(
         existing_persona.is_visible = is_visible
         existing_persona.search_start_date = search_start_date
         existing_persona.labels = labels or []
+        existing_persona.is_default_persona = is_default_persona
         # Do not delete any associations manually added unless
         # a new updated list is provided
         if document_sets is not None:
@@ -590,6 +599,23 @@ def delete_old_default_personas(
     db_session.commit()
 
 
+def update_persona_is_default(
+    persona_id: int,
+    is_default: bool,
+    db_session: Session,
+    user: User | None = None,
+) -> None:
+    persona = fetch_persona_by_id_for_user(
+        db_session=db_session, persona_id=persona_id, user=user, get_editable=True
+    )
+
+    if not persona.is_public:
+        persona.is_public = True
+
+    persona.is_default_persona = is_default
+    db_session.commit()
+
+
 def update_persona_visibility(
     persona_id: int,
     is_visible: bool,

backend/onyx/db/users.py

@@ -6,6 +6,7 @@ from fastapi import HTTPException
 from fastapi_users.password import PasswordHelper
 from sqlalchemy import func
 from sqlalchemy import select
+from sqlalchemy.exc import IntegrityError
 from sqlalchemy.orm import Session
 from sqlalchemy.sql import expression
 from sqlalchemy.sql.elements import ColumnElement
@@ -274,7 +275,7 @@ def _generate_ext_permissioned_user(email: str) -> User:
 
 
 def batch_add_ext_perm_user_if_not_exists(
-    db_session: Session, emails: list[str]
+    db_session: Session, emails: list[str], continue_on_error: bool = False
 ) -> list[User]:
     lower_emails = [email.lower() for email in emails]
     found_users, missing_lower_emails = _get_users_by_emails(db_session, lower_emails)
@@ -283,10 +284,23 @@ def batch_add_ext_perm_user_if_not_exists(
     for email in missing_lower_emails:
         new_users.append(_generate_ext_permissioned_user(email=email))
 
-    db_session.add_all(new_users)
-    db_session.commit()
-
-    return found_users + new_users
+    try:
+        db_session.add_all(new_users)
+        db_session.commit()
+    except IntegrityError:
+        db_session.rollback()
+        if not continue_on_error:
+            raise
+        for user in new_users:
+            try:
+                db_session.add(user)
+                db_session.commit()
+            except IntegrityError:
+                db_session.rollback()
+                continue
+    # Fetch all users again to ensure we have the most up-to-date list
+    all_users, _ = _get_users_by_emails(db_session, lower_emails)
+    return all_users
 
 
 def delete_user_from_db(

backend/onyx/document_index/vespa/index.py

@@ -17,6 +17,7 @@ from uuid import UUID
 
 import httpx  # type: ignore
 import requests  # type: ignore
+from retry import retry
 
 from onyx.configs.chat_configs import DOC_TIME_DECAY
 from onyx.configs.chat_configs import NUM_RETURNED_HITS
@@ -549,6 +550,11 @@ class VespaIndex(DocumentIndex):
             time.monotonic() - update_start,
         )
 
+    @retry(
+        tries=3,
+        delay=1,
+        backoff=2,
+    )
     def _update_single_chunk(
         self,
         doc_chunk_id: UUID,
@@ -559,6 +565,7 @@ class VespaIndex(DocumentIndex):
     ) -> None:
         """
         Update a single "chunk" (document) in Vespa using its chunk ID.
+        Retries if we encounter transient HTTPStatusError (e.g., overload).
         """
 
         update_dict: dict[str, dict] = {"fields": {}}
@@ -567,13 +574,11 @@ class VespaIndex(DocumentIndex):
             update_dict["fields"][BOOST] = {"assign": fields.boost}
 
         if fields.document_sets is not None:
-            # WeightedSet<string> needs a map { item: weight, ... }
             update_dict["fields"][DOCUMENT_SETS] = {
                 "assign": {document_set: 1 for document_set in fields.document_sets}
             }
 
         if fields.access is not None:
-            # Similar to above
             update_dict["fields"][ACCESS_CONTROL_LIST] = {
                 "assign": {acl_entry: 1 for acl_entry in fields.access.to_acl()}
             }
@@ -585,7 +590,10 @@ class VespaIndex(DocumentIndex):
             logger.error("Update request received but nothing to update.")
             return
 
-        vespa_url = f"{DOCUMENT_ID_ENDPOINT.format(index_name=index_name)}/{doc_chunk_id}?create=true"
+        vespa_url = (
+            f"{DOCUMENT_ID_ENDPOINT.format(index_name=index_name)}/{doc_chunk_id}"
+            "?create=true"
+        )
 
         try:
             resp = http_client.put(
@@ -595,8 +603,11 @@ class VespaIndex(DocumentIndex):
             )
             resp.raise_for_status()
         except httpx.HTTPStatusError as e:
-            error_message = f"Failed to update doc chunk {doc_chunk_id} (doc_id={doc_id}). Details: {e.response.text}"
-            logger.error(error_message)
+            logger.error(
+                f"Failed to update doc chunk {doc_chunk_id} (doc_id={doc_id}). "
+                f"Details: {e.response.text}"
+            )
+            # Re-raise so the @retry decorator will catch and retry
             raise
 
     def update_single(

backend/onyx/document_index/vespa/indexing_utils.py

@@ -146,6 +146,23 @@ def _index_vespa_chunk(
 
     title = document.get_title_for_document_index()
 
+    metadata_json = document.metadata
+    cleaned_metadata_json: dict[str, str | list[str]] = {}
+    for key, value in metadata_json.items():
+        cleaned_key = remove_invalid_unicode_chars(key)
+        if isinstance(value, list):
+            cleaned_metadata_json[cleaned_key] = [
+                remove_invalid_unicode_chars(item) for item in value
+            ]
+        else:
+            cleaned_metadata_json[cleaned_key] = remove_invalid_unicode_chars(value)
+
+    metadata_list = document.get_metadata_str_attributes()
+    if metadata_list:
+        metadata_list = [
+            remove_invalid_unicode_chars(metadata) for metadata in metadata_list
+        ]
+
     vespa_document_fields = {
         DOCUMENT_ID: document.id,
         CHUNK_ID: chunk.chunk_id,
@@ -166,10 +183,10 @@ def _index_vespa_chunk(
         SEMANTIC_IDENTIFIER: remove_invalid_unicode_chars(document.semantic_identifier),
         SECTION_CONTINUATION: chunk.section_continuation,
         LARGE_CHUNK_REFERENCE_IDS: chunk.large_chunk_reference_ids,
-        METADATA: json.dumps(document.metadata),
+        METADATA: json.dumps(cleaned_metadata_json),
         # Save as a list for efficient extraction as an Attribute
-        METADATA_LIST: chunk.source_document.get_metadata_str_attributes(),
-        METADATA_SUFFIX: chunk.metadata_suffix_keyword,
+        METADATA_LIST: metadata_list,
+        METADATA_SUFFIX: remove_invalid_unicode_chars(chunk.metadata_suffix_keyword),
         EMBEDDINGS: embeddings_name_vector_map,
         TITLE_EMBEDDING: chunk.title_embedding,
         DOC_UPDATED_AT: _vespa_get_updated_at_attribute(document.doc_updated_at),

backend/onyx/llm/chat_llm.py

@@ -27,6 +27,7 @@ from langchain_core.prompt_values import PromptValue
 
 from onyx.configs.app_configs import LOG_DANSWER_MODEL_INTERACTIONS
 from onyx.configs.app_configs import MOCK_LLM_RESPONSE
+from onyx.configs.chat_configs import QA_TIMEOUT
 from onyx.configs.model_configs import (
     DISABLE_LITELLM_STREAMING,
 )
@@ -35,6 +36,7 @@ from onyx.configs.model_configs import LITELLM_EXTRA_BODY
 from onyx.llm.interfaces import LLM
 from onyx.llm.interfaces import LLMConfig
 from onyx.llm.interfaces import ToolChoiceOptions
+from onyx.llm.utils import model_is_reasoning_model
 from onyx.server.utils import mask_string
 from onyx.utils.logger import setup_logger
 from onyx.utils.long_term_log import LongTermLogger
@@ -50,6 +52,18 @@ litellm.telemetry = False
 _LLM_PROMPT_LONG_TERM_LOG_CATEGORY = "llm_prompt"
 
 
+class LLMTimeoutError(Exception):
+    """
+    Exception raised when an LLM call times out.
+    """
+
+
+class LLMRateLimitError(Exception):
+    """
+    Exception raised when an LLM call is rate limited.
+    """
+
+
 def _base_msg_to_role(msg: BaseMessage) -> str:
     if isinstance(msg, HumanMessage) or isinstance(msg, HumanMessageChunk):
         return "user"
@@ -229,15 +243,15 @@ class DefaultMultiLLM(LLM):
     def __init__(
         self,
         api_key: str | None,
-        timeout: int,
         model_provider: str,
         model_name: str,
+        timeout: int | None = None,
         api_base: str | None = None,
         api_version: str | None = None,
         deployment_name: str | None = None,
         max_output_tokens: int | None = None,
         custom_llm_provider: str | None = None,
-        temperature: float = GEN_AI_TEMPERATURE,
+        temperature: float | None = None,
         custom_config: dict[str, str] | None = None,
         extra_headers: dict[str, str] | None = None,
         extra_body: dict | None = LITELLM_EXTRA_BODY,
@@ -245,9 +259,16 @@ class DefaultMultiLLM(LLM):
         long_term_logger: LongTermLogger | None = None,
     ):
         self._timeout = timeout
+        if timeout is None:
+            if model_is_reasoning_model(model_name):
+                self._timeout = QA_TIMEOUT * 10  # Reasoning models are slow
+            else:
+                self._timeout = QA_TIMEOUT
+
+        self._temperature = GEN_AI_TEMPERATURE if temperature is None else temperature
+
         self._model_provider = model_provider
         self._model_version = model_name
-        self._temperature = temperature
         self._api_key = api_key
         self._deployment_name = deployment_name
         self._api_base = api_base
@@ -380,6 +401,7 @@ class DefaultMultiLLM(LLM):
         tool_choice: ToolChoiceOptions | None,
         stream: bool,
         structured_response_format: dict | None = None,
+        timeout_override: int | None = None,
     ) -> litellm.ModelResponse | litellm.CustomStreamWrapper:
         # litellm doesn't accept LangChain BaseMessage objects, so we need to convert them
         # to a dict representation
@@ -387,9 +409,14 @@ class DefaultMultiLLM(LLM):
         self._record_call(processed_prompt)
 
         try:
+            print(
+                "model is",
+                f"{self.config.model_provider}/{self.config.deployment_name or self.config.model_name}",
+            )
             return litellm.completion(
                 mock_response=MOCK_LLM_RESPONSE,
                 # model choice
+                # model="openai/gpt-4",
                 model=f"{self.config.model_provider}/{self.config.deployment_name or self.config.model_name}",
                 # NOTE: have to pass in None instead of empty string for these
                 # otherwise litellm can have some issues with bedrock
@@ -405,7 +432,7 @@ class DefaultMultiLLM(LLM):
                 stream=stream,
                 # model params
                 temperature=0,
-                timeout=self._timeout,
+                timeout=timeout_override or self._timeout,
                 # For now, we don't support parallel tool calls
                 # NOTE: we can't pass this in if tools are not specified
                 # or else OpenAI throws an error
@@ -424,6 +451,12 @@ class DefaultMultiLLM(LLM):
         except Exception as e:
             self._record_error(processed_prompt, e)
             # for break pointing
+            if isinstance(e, litellm.Timeout):
+                raise LLMTimeoutError(e)
+
+            elif isinstance(e, litellm.RateLimitError):
+                raise LLMRateLimitError(e)
+
             raise e
 
     @property
@@ -444,6 +477,7 @@ class DefaultMultiLLM(LLM):
         tools: list[dict] | None = None,
         tool_choice: ToolChoiceOptions | None = None,
         structured_response_format: dict | None = None,
+        timeout_override: int | None = None,
     ) -> BaseMessage:
         if LOG_DANSWER_MODEL_INTERACTIONS:
             self.log_model_configs()
@@ -451,7 +485,12 @@ class DefaultMultiLLM(LLM):
         response = cast(
             litellm.ModelResponse,
             self._completion(
-                prompt, tools, tool_choice, False, structured_response_format
+                prompt=prompt,
+                tools=tools,
+                tool_choice=tool_choice,
+                stream=False,
+                structured_response_format=structured_response_format,
+                timeout_override=timeout_override,
             ),
         )
         choice = response.choices[0]
@@ -469,19 +508,31 @@ class DefaultMultiLLM(LLM):
         tools: list[dict] | None = None,
         tool_choice: ToolChoiceOptions | None = None,
         structured_response_format: dict | None = None,
+        timeout_override: int | None = None,
     ) -> Iterator[BaseMessage]:
         if LOG_DANSWER_MODEL_INTERACTIONS:
             self.log_model_configs()
 
         if DISABLE_LITELLM_STREAMING:
-            yield self.invoke(prompt, tools, tool_choice, structured_response_format)
+            yield self.invoke(
+                prompt,
+                tools,
+                tool_choice,
+                structured_response_format,
+                timeout_override,
+            )
             return
 
         output = None
         response = cast(
             litellm.CustomStreamWrapper,
             self._completion(
-                prompt, tools, tool_choice, True, structured_response_format
+                prompt=prompt,
+                tools=tools,
+                tool_choice=tool_choice,
+                stream=True,
+                structured_response_format=structured_response_format,
+                timeout_override=timeout_override,
             ),
         )
         try:

backend/onyx/llm/custom_llm.py

@@ -81,6 +81,7 @@ class CustomModelServer(LLM):
         tools: list[dict] | None = None,
         tool_choice: ToolChoiceOptions | None = None,
         structured_response_format: dict | None = None,
+        timeout_override: int | None = None,
     ) -> BaseMessage:
         return self._execute(prompt)
 
@@ -90,5 +91,6 @@ class CustomModelServer(LLM):
         tools: list[dict] | None = None,
         tool_choice: ToolChoiceOptions | None = None,
         structured_response_format: dict | None = None,
+        timeout_override: int | None = None,
     ) -> Iterator[BaseMessage]:
         yield self._execute(prompt)

backend/onyx/llm/factory.py

@@ -2,7 +2,6 @@ from typing import Any
 
 from onyx.chat.models import PersonaOverrideConfig
 from onyx.configs.app_configs import DISABLE_GENERATIVE_AI
-from onyx.configs.chat_configs import QA_TIMEOUT
 from onyx.configs.model_configs import GEN_AI_MODEL_FALLBACK_MAX_TOKENS
 from onyx.configs.model_configs import GEN_AI_TEMPERATURE
 from onyx.db.engine import get_session_context_manager
@@ -88,8 +87,8 @@ def get_llms_for_persona(
 
 
 def get_default_llms(
-    timeout: int = QA_TIMEOUT,
-    temperature: float = GEN_AI_TEMPERATURE,
+    timeout: int | None = None,
+    temperature: float | None = None,
     additional_headers: dict[str, str] | None = None,
     long_term_logger: LongTermLogger | None = None,
 ) -> tuple[LLM, LLM]:
@@ -138,7 +137,7 @@ def get_llm(
     api_version: str | None = None,
     custom_config: dict[str, str] | None = None,
     temperature: float | None = None,
-    timeout: int = QA_TIMEOUT,
+    timeout: int | None = None,
     additional_headers: dict[str, str] | None = None,
     long_term_logger: LongTermLogger | None = None,
 ) -> LLM:

backend/onyx/llm/interfaces.py

@@ -90,12 +90,13 @@ class LLM(abc.ABC):
         tools: list[dict] | None = None,
         tool_choice: ToolChoiceOptions | None = None,
         structured_response_format: dict | None = None,
+        timeout_override: int | None = None,
     ) -> BaseMessage:
         self._precall(prompt)
         # TODO add a postcall to log model outputs independent of concrete class
         # implementation
         return self._invoke_implementation(
-            prompt, tools, tool_choice, structured_response_format
+            prompt, tools, tool_choice, structured_response_format, timeout_override
         )
 
     @abc.abstractmethod
@@ -105,6 +106,7 @@ class LLM(abc.ABC):
         tools: list[dict] | None = None,
         tool_choice: ToolChoiceOptions | None = None,
         structured_response_format: dict | None = None,
+        timeout_override: int | None = None,
     ) -> BaseMessage:
         raise NotImplementedError
 
@@ -114,12 +116,13 @@ class LLM(abc.ABC):
         tools: list[dict] | None = None,
         tool_choice: ToolChoiceOptions | None = None,
         structured_response_format: dict | None = None,
+        timeout_override: int | None = None,
     ) -> Iterator[BaseMessage]:
         self._precall(prompt)
         # TODO add a postcall to log model outputs independent of concrete class
         # implementation
         messages = self._stream_implementation(
-            prompt, tools, tool_choice, structured_response_format
+            prompt, tools, tool_choice, structured_response_format, timeout_override
         )
 
         tokens = []
@@ -138,5 +141,6 @@ class LLM(abc.ABC):
         tools: list[dict] | None = None,
         tool_choice: ToolChoiceOptions | None = None,
         structured_response_format: dict | None = None,
+        timeout_override: int | None = None,
     ) -> Iterator[BaseMessage]:
         raise NotImplementedError

backend/onyx/llm/llm_provider_options.py

@@ -29,11 +29,11 @@ OPENAI_PROVIDER_NAME = "openai"
 OPEN_AI_MODEL_NAMES = [
     "o3-mini",
     "o1-mini",
-    "o1-preview",
-    "o1-2024-12-17",
+    "o1",
     "gpt-4",
     "gpt-4o",
     "gpt-4o-mini",
+    "o1-preview",
     "gpt-4-turbo",
     "gpt-4-turbo-preview",
     "gpt-4-1106-preview",

backend/onyx/llm/utils.py

@@ -543,3 +543,14 @@ def model_supports_image_input(model_name: str, model_provider: str) -> bool:
             f"Failed to get model object for {model_provider}/{model_name}"
         )
         return False
+
+
+def model_is_reasoning_model(model_name: str) -> bool:
+    _REASONING_MODEL_NAMES = [
+        "o1",
+        "o1-mini",
+        "o3-mini",
+        "deepseek-reasoner",
+        "deepseek-r1",
+    ]
+    return model_name.lower() in _REASONING_MODEL_NAMES

backend/onyx/main.py

@@ -238,14 +238,20 @@ async def lifespan(app: FastAPI) -> AsyncGenerator[None, None]:
         await close_auth_limiter()
 
 
-def log_http_error(_: Request, exc: Exception) -> JSONResponse:
+def log_http_error(request: Request, exc: Exception) -> JSONResponse:
     status_code = getattr(exc, "status_code", 500)
 
     if isinstance(exc, BasicAuthenticationError):
-        # For BasicAuthenticationError, just log a brief message without stack trace (almost always spam)
-        logger.warning(f"Authentication failed: {str(exc)}")
+        # For BasicAuthenticationError, just log a brief message without stack trace
+        # (almost always spammy)
+        logger.debug(f"Authentication failed: {str(exc)}")
+
+    elif status_code == 404 and request.url.path == "/metrics":
+        # Log 404 errors for the /metrics endpoint with debug level
+        logger.debug(f"404 error for /metrics endpoint: {str(exc)}")
 
     elif status_code >= 400:
+        print("FORMATTING ERROR")
         error_msg = f"{str(exc)}\n"
         error_msg += "".join(traceback.format_tb(exc.__traceback__))
         logger.error(error_msg)

backend/onyx/natural_language_processing/utils.py

@@ -99,7 +99,7 @@ def _check_tokenizer_cache(
 
         if not tokenizer:
             logger.info(
-                f"Falling back to default embedding model: {DOCUMENT_ENCODER_MODEL}"
+                f"Falling back to default embedding model tokenizer: {DOCUMENT_ENCODER_MODEL}"
             )
             tokenizer = HuggingFaceTokenizer(DOCUMENT_ENCODER_MODEL)
 

backend/onyx/prompts/agent_search.py

@@ -1,3 +1,7 @@
+from onyx.agents.agent_search.shared_graph_utils.constants import (
+    AGENT_ANSWER_SEPARATOR,
+)
+
 # Standards
 SEPARATOR_LINE = "-------"
 SEPARATOR_LINE_LONG = "---------------"
@@ -5,8 +9,6 @@ UNKNOWN_ANSWER = "I do not have enough information to answer this question."
 NO_RECOVERED_DOCS = "No relevant information recovered"
 YES = "yes"
 NO = "no"
-
-
 # Framing/Support/Template Prompts
 HISTORY_FRAMING_PROMPT = f"""
 For more context, here is the history of the conversation so far that preceded this question:
@@ -16,6 +18,43 @@ For more context, here is the history of the conversation so far that preceded t
 """.strip()
 
 
+COMMON_RAG_RULES = f"""
+IMPORTANT RULES:
+ - If you cannot reliably answer the question solely using the provided information, say that you cannot reliably answer. \
+You may give some additional facts you learned, but do not try to invent an answer.
+
+ - If the information is empty or irrelevant, just say "{UNKNOWN_ANSWER}".
+
+ - If the information is relevant but not fully conclusive, provide an answer to the extent you can but also specify that \
+the information is not conclusive and why.
+
+- When constructing/considering categories, focus less on the question and more on the context actually provided! \
+Example: if the question is about the products of company A, and the content provided lists a number of products, \
+do automatically NOT ASSUME that those belong to company A!  So you cannot list those as products of company A, despite \
+the fact that the question is about company A's products. What you should say instead is maybe something like \
+"Here are a number of products, but I cannot say whether some or all of them belong to company A: \
+<proceed with listing the products>". It is ABSOLUTELY ESSENTIAL that the answer constructed reflects \
+actual knowledge. For that matter, also consider the title of the document and other information that may be \
+provided. If that does not make it clear that - in the example above - the products belong to company A, \
+then do not list them as products of company A, just maybe as "A list products that may not necessarily \
+belong to company A". THIS IS IMPORTANT!
+
+- Related, if the context provides a list of items with associated data or other information that seems \
+to align with the categories in the question, but does not specify whether the items or the information is \
+specific to the exact requested category, then present the information with a disclaimer. Use a title such as \
+"I am not sure whether these items (or the information provided) is specific to [relevant category] or whether \
+these are all [specific group], but I found this information may be helpful:" \
+followed by the list of items and associated data/or information discovered.
+
+ - Do not group together items amongst one headline where not all items belong to the category of the headline! \
+(Example: "Products used by Company A" where some products listed are not built by Company A, but other companies,
+or it is not clear that the products are built by Company A). Only state what you know for sure!
+
+ - Do NOT perform any calculations in the answer! Just report on facts.
+
+ - If appropriate, organizing your answer in bullet points is often useful.
+""".strip()
+
 ASSISTANT_SYSTEM_PROMPT_DEFAULT = "You are an assistant for question-answering tasks."
 
 ASSISTANT_SYSTEM_PROMPT_PERSONA = f"""
@@ -129,20 +168,44 @@ History summary:
 # Sub-question
 # Intentionally left a copy in case we want to modify this one differently
 INITIAL_QUESTION_DECOMPOSITION_PROMPT = f"""
-Decompose the initial user question into no more than 3 appropriate sub-questions that help to answer the \
-original question. The purpose for this decomposition may be to:
-  1) isolate individual entities (i.e., 'compare sales of company A and company B' -> \
+Please create a list of no more than 3 sub-questions whose answers would help to inform the answer \
+to the initial question.
+
+The purpose for these sub-questions could be:
+  1) decomposition to isolate individual entities (i.e., 'compare sales of company A and company B' -> \
 ['what are sales for company A', 'what are sales for company B'])
-  2) clarify or disambiguate ambiguous terms (i.e., 'what is our success with company A' -> \
+
+  2) clarification and/or disambiguation of ambiguous terms (i.e., 'what is our success with company A' -> \
 ['what are our sales with company A','what is our market share with company A', \
 'is company A a reference customer for us', etc.])
-  3) if a term or a metric is essentially clear, but it could relate to various components of an entity and you \
-are generally familiar with the entity, then you can decompose the question into sub-questions that are more \
-specific to components (i.e., 'what do we do to improve scalability of product X', 'what do we to to improve \
-scalability of product X', 'what do we do to improve stability of product X', ...])
-  4) research an area that could really help to answer the question.
 
-Here is the initial question to decompose:
+  3) if a term or a metric is essentially clear, but it could relate to various aspects of an entity and you \
+are generally familiar with the entity, then you can create sub-questions that are more \
+specific (i.e.,  'what do we do to improve product X' -> 'what do we do to improve scalability of product X', \
+'what do we do to improve performance of product X', 'what do we do to improve stability of product X', ...)
+
+  4) research individual questions and areas that should really help to ultimately answer the question.
+
+Important:
+
+ - Each sub-question should lend itself to be answered by a RAG system. Correspondingly, phrase the question \
+in a way that is amenable to that. An example set of sub-questions based on an initial question could look like this:
+'what can I do to improve the performance of workflow X' -> \
+'what are the settings affecting performance for workflow X', 'are there complaints and bugs related to \
+workflow X performance', 'what are performance benchmarks for workflow X', ...
+
+ - Consequently, again, don't just decompose, but make sure that the sub-questions have the proper form. I.e., no \
+ 'I', etc.
+
+ - Do not(!) create sub-questions that are clarifying question to the person who asked the question, \
+like making suggestions or asking the user for more information! This is not useful for the actual \
+question-answering process! You need to take the information from the user as it is given to you! \
+For example, should the question be of the type 'why does product X perform poorly for customer A', DO NOT create a \
+sub-question of the type 'what are the settings that customer A uses for product X?'! A valid sub-question \
+could rather be 'which settings for product X have been shown to lead to poor performance for customers?'
+
+
+And here is the initial question to create sub-questions for, so that you have the full context:
 {SEPARATOR_LINE}
 {{question}}
 {SEPARATOR_LINE}
@@ -150,7 +213,79 @@ Here is the initial question to decompose:
 {{history}}
 
 Do NOT include any text in your answer outside of the list of sub-questions!
-Please formulate your answer as a newline-separated list of questions like so:
+Please formulate your answer as a newline-separated list of questions like so (and please ONLY ANSWER WITH THIS LIST! Do not \
+add any explanations or other text!):
+
+ <sub-question>
+ <sub-question>
+ <sub-question>
+ ...
+
+Answer:
+""".strip()
+
+# INITIAL PHASE - AWARE OF REFINEMENT
+# Sub-question
+# Suggest augmenting question generation as well, that a future refinement phase could use
+# to generate new questions
+# Intentionally left a copy in case we want to modify this one differently
+INITIAL_QUESTION_DECOMPOSITION_PROMPT_ASSUMING_REFINEMENT = f"""
+Please create a list of no more than 3 sub-questions whose answers would help to inform the answer \
+to the initial question.
+
+The purpose for these sub-questions could be:
+  1) decomposition to isolate individual entities (i.e., 'compare sales of company A and company B' -> \
+['what are sales for company A', 'what are sales for company B'])
+
+  2) clarification and/or disambiguation of ambiguous terms (i.e., 'what is our success with company A' -> \
+['what are our sales with company A','what is our market share with company A', \
+'is company A a reference customer for us', etc.])
+
+  3) if a term or a metric is essentially clear, but it could relate to various aspects of an entity and you \
+are generally familiar with the entity, then you can create sub-questions that are more \
+specific (i.e.,  'what do we do to improve product X' -> 'what do we do to improve scalability of product X', \
+'what do we do to improve performance of product X', 'what do we do to improve stability of product X', ...)
+
+  4) research individual questions and areas that should really help to ultimately answer the question.
+
+  5) if meaningful, find relevant facts that may inform another set of sub-questions generate after the set you \
+create now are answered. Example: 'which products have we implemented at company A, and is this different to \
+its competitors?'  could potentially create sub-questions 'what products have we implemented at company A', \
+and 'who are the competitors of company A'. The additional round of sub-question generation which sees the \
+answers for this initial round of sub-question creation could then use the answer to the second sub-question \
+(which could be 'company B and C are competitors of company A') to then ask 'which products have we implemented \
+at company B', 'which products have we implemented at company C'...
+
+Important:
+
+ - Each sub-question should lend itself to be answered by a RAG system. Correspondingly, phrase the question \
+in a way that is amenable to that. An example set of sub-questions based on an initial question could look like this:
+'what can I do to improve the performance of workflow X' -> \
+'what are the settings affecting performance for workflow X', 'are there complaints and bugs related to \
+workflow X performance', 'what are performance benchmarks for workflow X', ...
+
+ - Consequently, again, don't just decompose, but make sure that the sub-questions have the proper form. I.e., no \
+ 'I', etc.
+
+ - Do not(!) create sub-questions that are clarifying question to the person who asked the question, \
+like making suggestions or asking the user for more information! This is not useful for the actual \
+question-answering process! You need to take the information from the user as it is given to you! \
+For example, should the question be of the type 'why does product X perform poorly for customer A', DO NOT create a \
+sub-question of the type 'what are the settings that customer A uses for product X?'! A valid sub-question \
+could rather be 'which settings for product X have been shown to lead to poor performance for customers?'
+
+
+And here is the initial question to create sub-questions for:
+{SEPARATOR_LINE}
+{{question}}
+{SEPARATOR_LINE}
+
+{{history}}
+
+Do NOT include any text in your answer outside of the list of sub-questions!
+Please formulate your answer as a newline-separated list of questions like so (and please ONLY ANSWER WITH THIS LIST! Do not \
+add any explanations or other text!):
+
  <sub-question>
  <sub-question>
  <sub-question>
@@ -162,23 +297,47 @@ Answer:
 
 # TODO: combine shared pieces with INITIAL_QUESTION_DECOMPOSITION_PROMPT
 INITIAL_DECOMPOSITION_PROMPT_QUESTIONS_AFTER_SEARCH = f"""
-Decompose the initial user question into no more than 3 appropriate sub-questions that help to answer the \
-original question. The purpose for this decomposition may be to:
-  1) isolate individual entities (i.e., 'compare sales of company A and company B' -> \
+Please create a list of no more than 3 sub-questions whose answers would help to inform the answer \
+to the initial question.
+
+The purpose for these sub-questions could be:
+  1) decomposition to isolate individual entities (i.e., 'compare sales of company A and company B' -> \
 ['what are sales for company A', 'what are sales for company B'])
-  2) clarify or disambiguate ambiguous terms (i.e., 'what is our success with company A' -> \
+
+  2) clarification and/or disambiguation of ambiguous terms (i.e., 'what is our success with company A' -> \
 ['what are our sales with company A','what is our market share with company A', \
 'is company A a reference customer for us', etc.])
-  3) if a term or a metric is essentially clear, but it could relate to various components of an entity and you \
-are generally familiar with the entity, then you can decompose the question into sub-questions that are more \
-specific to components (i.e., 'what do we do to improve scalability of product X', 'what do we to to improve \
-scalability of product X', 'what do we do to improve stability of product X', ...])
-  4) research an area that could really help to answer the question.
+
+  3) if a term or a metric is essentially clear, but it could relate to various aspects of an entity and you \
+are generally familiar with the entity, then you can create sub-questions that are more \
+specific (i.e.,  'what do we do to improve product X' -> 'what do we do to improve scalability of product X', \
+'what do we do to improve performance of product X', 'what do we do to improve stability of product X', ...)
+
+  4) research individual questions and areas that should really help to ultimately answer the question.
+
+Important:
+
+ - Each sub-question should lend itself to be answered by a RAG system. Correspondingly, phrase the question \
+in a way that is amenable to that. An example set of sub-questions based on an initial question could look like this:
+'what can I do to improve the performance of workflow X' -> \
+'what are the settings affecting performance for workflow X', 'are there complaints and bugs related to \
+workflow X performance', 'what are performance benchmarks for workflow X', ...
+
+ - Consequently, again, don't just decompose, but make sure that the sub-questions have the proper form. I.e., no \
+ 'I', etc.
+
+ - Do not(!) create sub-questions that are clarifying question to the person who asked the question, \
+like making suggestions or asking the user for more information! This is not useful for the actual \
+question-answering process! You need to take the information from the user as it is given to you! \
+For example, should the question be of the type 'why does product X perform poorly for customer A', DO NOT create a \
+sub-question of the type 'what are the settings that customer A uses for product X?'! A valid sub-question \
+could rather be 'which settings for product X have been shown to lead to poor performance for customers?'
+
 
 To give you some context, you will see below also some documents that may relate to the question. Please only \
 use this information to learn what the question is approximately asking about, but do not focus on the details \
 to construct the sub-questions! Also, some of the entities, relationships and terms that are in the dataset may \
-not be in these few documents, so DO NOT focussed too much on the documents when constructing the sub-questions! \
+not be in these few documents, so DO NOT focus too much on the documents when constructing the sub-questions! \
 Decomposition and disambiguations are most important!
 
 Here are the sample docs to give you some context:
@@ -186,7 +345,7 @@ Here are the sample docs to give you some context:
 {{sample_doc_str}}
 {SEPARATOR_LINE}
 
-And here is the initial question to decompose:
+And here is the initial question to create sub-questions for, so that you have the full context:
 {SEPARATOR_LINE}
 {{question}}
 {SEPARATOR_LINE}
@@ -194,7 +353,9 @@ And here is the initial question to decompose:
 {{history}}
 
 Do NOT include any text in your answer outside of the list of sub-questions!\
-Please formulate your answer as a newline-separated list of questions like so:
+Please formulate your answer as a newline-separated list of questions like so (and please ONLY ANSWER WITH THIS LIST! Do not \
+add any explanations or other text!):
+
  <sub-question>
  <sub-question>
  <sub-question>
@@ -203,6 +364,84 @@ Please formulate your answer as a newline-separated list of questions like so:
 Answer:
 """.strip()
 
+INITIAL_DECOMPOSITION_PROMPT_QUESTIONS_AFTER_SEARCH_ASSUMING_REFINEMENT = f"""
+Please create a list of no more than 3 sub-questions whose answers would help to inform the answer \
+to the initial question.
+
+The purpose for these sub-questions could be:
+  1) decomposition to isolate individual entities (i.e., 'compare sales of company A and company B' -> \
+['what are sales for company A', 'what are sales for company B'])
+
+  2) clarification and/or disambiguation of ambiguous terms (i.e., 'what is our success with company A' -> \
+['what are our sales with company A','what is our market share with company A', \
+'is company A a reference customer for us', etc.])
+
+  3) if a term or a metric is essentially clear, but it could relate to various aspects of an entity and you \
+are generally familiar with the entity, then you can create sub-questions that are more \
+specific (i.e.,  'what do we do to improve product X' -> 'what do we do to improve scalability of product X', \
+'what do we do to improve performance of product X', 'what do we do to improve stability of product X', ...)
+
+  4) research individual questions and areas that should really help to ultimately answer the question.
+
+  5) if applicable and useful, consider using sub-questions to gather relevant information that can inform a \
+subsequent set of sub-questions. The answers to your initial sub-questions will be available when generating \
+the next set.
+For example, if you start with the question, "Which products have we implemented at Company A, and how does \
+this compare to its competitors?" you might first create sub-questions like "What products have we implemented \
+at Company A?" and "Who are the competitors of Company A?"
+The answer to the second sub-question, such as "Company B and C are competitors of Company A," can then be used \
+to generate more specific sub-questions in the next round, like "Which products have we implemented at Company B?" \
+and "Which products have we implemented at Company C?"
+
+You'll be the judge!
+
+Important:
+
+ - Each sub-question should lend itself to be answered by a RAG system. Correspondingly, phrase the question \
+in a way that is amenable to that. An example set of sub-questions based on an initial question could look like this:
+'what can I do to improve the performance of workflow X' -> \
+'what are the settings affecting performance for workflow X', 'are there complaints and bugs related to \
+workflow X performance', 'what are performance benchmarks for workflow X', ...
+
+ - Consequently, again, don't just decompose, but make sure that the sub-questions have the proper form. I.e., no \
+ 'I', etc.
+
+ - Do not(!) create sub-questions that are clarifying question to the person who asked the question, \
+like making suggestions or asking the user for more information! This is not useful for the actual \
+question-answering process! You need to take the information from the user as it is given to you! \
+For example, should the question be of the type 'why does product X perform poorly for customer A', DO NOT create a \
+sub-question of the type 'what are the settings that customer A uses for product X?'! A valid sub-question \
+could rather be 'which settings for product X have been shown to lead to poor performance for customers?'
+
+To give you some context, you will see below also some documents that may relate to the question. Please only \
+use this information to learn what the question is approximately asking about, but do not focus on the details \
+to construct the sub-questions! Also, some of the entities, relationships and terms that are in the dataset may \
+not be in these few documents, so DO NOT focus too much on the documents when constructing the sub-questions! \
+Decomposition and disambiguations are most important!
+
+Here are the sample docs to give you some context:
+{SEPARATOR_LINE}
+{{sample_doc_str}}
+{SEPARATOR_LINE}
+
+And here is the initial question to create sub-questions for, so that you have the full context:
+{SEPARATOR_LINE}
+{{question}}
+{SEPARATOR_LINE}
+
+{{history}}
+
+Do NOT include any text in your answer outside of the list of sub-questions!\
+Please formulate your answer as a newline-separated list of questions like so (and please ONLY ANSWER WITH THIS LIST! Do not \
+add any explanations or other text!):
+
+ <sub-question>
+ <sub-question>
+ <sub-question>
+ ...
+
+Answer:
+""".strip()
 
 # Retrieval
 QUERY_REWRITING_PROMPT = f"""
@@ -257,23 +496,35 @@ Answer:
 """.strip()
 
 
-# Sub-Question Anser Generation
+# Sub-Question Answer Generation
 SUB_QUESTION_RAG_PROMPT = f"""
 Use the context provided below - and only the provided context - to answer the given question. \
 (Note that the answer is in service of answering a broader question, given below as 'motivation').
 
-Again, only use the provided context and do not use your internal knowledge! If you cannot answer the \
-question based on the context, say "{UNKNOWN_ANSWER}". It is a matter of life and death that you do NOT \
-use your internal knowledge, just the provided information!
-
-Make sure that you keep all relevant information, specifically as it concerns to the ultimate goal. \
+Make sure that you keep all relevant information, specifically as it concerns the ultimate goal. \
 (But keep other details as well.)
 
-It is critical that you provide inline citations in the format [D1], [D2], [D3], etc! \
+{COMMON_RAG_RULES}
+
+ - Make sure that you only state what you actually can positively learn from the provided context! Particularly \
+don't make assumptions!  Example: if i) a question you should answer is asking for products of companies that \
+are competitors of company A, and ii) the context mentions products of companies A, B, C, D, E, etc., do NOT assume \
+that B, C, D, E, etc. are competitors of A! All you know is that these are products of a number of companies, and you \
+would have to rely on another question - that you do not have access to - to learn which companies are competitors of A.
+Correspondingly, you should not say that these are the products of competitors of A, but rather something like \
+"Here are some products of various companies".
+
+It is critical that you provide inline citations in the format [D1], [D2], [D3], etc! Please use format [D1][D2] and NOT \
+[D1, D2] format if you cite two or more documents together! \
 It is important that the citation is close to the information it supports. \
 Proper citations are very important to the user!
 
-For your general information, here is the ultimate motivation:
+Here is the document context for you to consider:
+{SEPARATOR_LINE}
+{{context}}
+{SEPARATOR_LINE}
+
+For your general information, here is the ultimate motivation for the question you need to answer:
 {SEPARATOR_LINE}
 {{original_question}}
 {SEPARATOR_LINE}
@@ -283,12 +534,8 @@ And here is the actual question I want you to answer based on the context above
 {{question}}
 {SEPARATOR_LINE}
 
-Here is the context:
-{SEPARATOR_LINE}
-{{context}}
-{SEPARATOR_LINE}
-
-Please keep your answer brief and concise, and focus on facts and data.
+Please keep your answer brief and concise, and focus on facts and data. (Again, only state what you see in the documents \
+for sure and communicate if/in what way this may or may not relate to the question you need to answer!)
 
 Answer:
 """.strip()
@@ -321,22 +568,18 @@ Use the information provided below - and only the provided information - to answ
 
 The information provided below consists of:
   1) a number of answered sub-questions - these are very important to help you organize your thoughts and your answer
-  2) a number of documents that deemed relevant for the question.
+  2) a number of documents that are deemed relevant for the question.
 
 {{history}}
 
-It is critical that you provide prover inline citations to documents in the format [D1], [D2], [D3], etc.! \
+It is critical that you provide proper inline citations to documents in the format [D1], [D2], [D3], etc.! \
 It is important that the citation is close to the information it supports. If you have multiple citations that support \
 a fact, please cite for example as [D1][D3], or [D2][D4], etc. \
 Feel free to also cite sub-questions in addition to documents, but make sure that you have documents cited with the \
 sub-question citation. If you want to cite both a document and a sub-question, please use [D1][Q3], or [D2][D7][Q4], etc. \
 Again, please NEVER cite sub-questions without a document citation! Proper citations are very important for the user!
 
-IMPORTANT RULES:
- - If you cannot reliably answer the question solely using the provided information, say that you cannot reliably answer. \
-You may give some additional facts you learned, but do not try to invent an answer.
- - If the information is empty or irrelevant, just say "{UNKNOWN_ANSWER}".
- - If the information is relevant but not fully conclusive, specify that the information is not conclusive and say why.
+{COMMON_RAG_RULES}
 
 Again, you should be sure that the answer is supported by the information provided!
 
@@ -361,7 +604,9 @@ And here is the question I want you to answer based on the information above:
 {{question}}
 {SEPARATOR_LINE}
 
-Please keep your answer brief and concise, and focus on facts and data.
+Please keep your answer brief and concise, and focus on facts and data. (Again, only state what you see in the documents for \
+sure and communicate if/in what way this may or may not relate to the question you need to answer! Use the answered \
+sub-questions as well, but be cautious and reconsider the docments again for validation.)
 
 Answer:
 """.strip()
@@ -376,11 +621,7 @@ The information provided below consists of a number of documents that were deeme
 
 {{history}}
 
-IMPORTANT RULES:
- - If you cannot reliably answer the question solely using the provided information, say that you cannot reliably answer. \
-You may give some additional facts you learned, but do not try to invent an answer.
- - If the information is irrelevant, just say "{UNKNOWN_ANSWER}".
- - If the information is relevant but not fully conclusive, specify that the information is not conclusive and say why.
+{COMMON_RAG_RULES}
 
 Again, you should be sure that the answer is supported by the information provided!
 
@@ -399,7 +640,8 @@ And here is the question I want you to answer based on the context above:
 {{question}}
 {SEPARATOR_LINE}
 
-Please keep your answer brief and concise, and focus on facts and data.
+Please keep your answer brief and concise, and focus on facts and data. (Again, only state what you see in the documents \
+for sure and communicate if/in what way this may or may not relate to the question you need to answer!)
 
 Answer:
 """.strip()
@@ -439,6 +681,12 @@ independently without the original question available
 - For each sub-question, please also provide a search term that can be used to retrieve relevant documents from a document store.
 - Consider specifically the sub-questions that were suggested but not answered. This is a sign that they are not answerable \
 with the available context, and you should not ask similar questions.
+ - Do not(!) create sub-questions that are clarifying question to the person who asked the question, \
+like making suggestions or asking the user for more information! This is not useful for the actual \
+question-answering process! You need to take the information from the user as it is given to you! \
+For example, should the question be of the type 'why does product X perform poorly for customer A', DO NOT create a \
+sub-question of the type 'what are the settings that customer A uses for product X?'! A valid sub-question \
+could rather be 'which settings for product X have been shown to lead to poor performance for customers?'
 
 Here is the initial question:
 {SEPARATOR_LINE}
@@ -474,7 +722,111 @@ objects/relationships/terms you can ask about! Do not ask about entities, terms
 Again, please find questions that are NOT overlapping too much with the already answered sub-questions or those that \
 already were suggested and failed. In other words - what can we try in addition to what has been tried so far?
 
-Generate the list of questions separated by one new line like this:
+Generate the list of questions separated by one new line like this (and please ONLY ANSWER WITH THIS LIST! Do not \
+add any explanations or other text!):
+
+<sub-question 1>
+<sub-question 2>
+<sub-question 3>
+...""".strip()
+
+REFINEMENT_QUESTION_DECOMPOSITION_PROMPT_W_INITIAL_SUBQUESTION_ANSWERS = f"""
+An initial user question needs to be answered. An initial answer has been provided but it wasn't quite good enough. \
+Also, some sub-questions had been answered and this information has been used to provide the initial answer. \
+Some other subquestions may have been suggested based on little knowledge, but they were not directly answerable. \
+Also, some entities, relationships and terms are given to you so that you have an idea of how the available data looks like.
+
+Your role is to generate 2-4 new sub-questions that would help to answer the initial question, considering:
+
+1) The initial question
+2) The initial answer that was found to be unsatisfactory
+3) The sub-questions that were answered AND their answers
+4) The sub-questions that were suggested but not answered (and that you should not repeat!)
+5) The entities, relationships and terms that were extracted from the context
+
+The individual questions should be answerable by a good RAG system. So a good idea would be to use the sub-questions to \
+resolve ambiguities and/or to separate the question for different entities that may be involved in the original question, \
+but in a way that does not duplicate questions that were already tried.
+
+Additional Guidelines:
+
+- The new sub-questions should be specific to the question and provide richer context for the question, resolve ambiguities, \
+or address shortcoming of the initial answer
+
+- Each new sub-question - when answered - should be relevant for the answer to the original question
+
+- The new sub-questions should be free from comparisons, ambiguities,judgements, aggregations, or any other complications that \
+may require extra context
+
+- The new sub-questions MUST have the full context of the original question so that it can be executed by a RAG system \
+independently without the original question available
+    Example:
+    - initial question: "What is the capital of France?"
+    - bad sub-question: "What is the name of the river there?"
+    - good sub-question: "What is the name of the river that flows through Paris?"
+
+    - For each new sub-question, please also provide a search term that can be used to retrieve relevant documents \
+from a document store.
+
+- Consider specifically the sub-questions that were suggested but not answered. This is a sign that they are not answerable \
+with the available context, and you should not ask similar questions.
+
+- Pay attention to the answers of previous sub-question to make your sub-questions more specific! \
+Often the initial sub-questions were set up to give you critical information that you should use to generate new sub-questions.\
+For example, if the answer to a an earlier sub-question is \
+'Company B and C are competitors of Company A', you should not ask now a new sub-question involving the term 'competitors', \
+as you already have the information to create a more precise question - you should instead explicitly reference \
+'Company B' and 'Company C' in your new sub-questions, as these are the competitors based on the previously answered question.
+
+- Be precise(!) and don't make inferences you cannot be sure about! For example, in the previous example \
+where Company B and Company C were identified as competitors of Company A, and then you also get information on \
+companies D and E, do not make the inference that these are also competitors of Company A! Stick to the information you have!
+(Also, don't assume that companies B and C arethe only competitors of A, unless stated!)
+
+- Do not(!) create sub-questions that are clarifying question *to the person who asked the question*, \
+like making suggestions or asking the user for more information! This is not useful for the actual \
+question-answering process! You need to take the information from the user as it is given to you! \
+For example, should the question be of the type 'why does product X perform poorly for customer A', DO NOT create a \
+sub-question of the type 'what are the settings that customer A uses for product X?'! A valid sub-question \
+could rather be 'which settings for product X have been shown to lead to poor performance for customers?'
+
+Here is the initial question:
+{SEPARATOR_LINE}
+{{question}}
+{SEPARATOR_LINE}
+{{history}}
+
+Here is the initial sub-optimal answer:
+{SEPARATOR_LINE}
+{{base_answer}}
+{SEPARATOR_LINE}
+
+Here are the sub-questions that were answered:
+{SEPARATOR_LINE}
+{{answered_subquestions_with_answers}}
+{SEPARATOR_LINE}
+
+Here are the sub-questions that were suggested but not answered:
+{SEPARATOR_LINE}
+{{failed_sub_questions}}
+{SEPARATOR_LINE}
+
+And here are the entities, relationships and terms extracted from the context:
+{SEPARATOR_LINE}
+{{entity_term_extraction_str}}
+{SEPARATOR_LINE}
+
+Please generate the list of good, fully contextualized sub-questions that would help to address the main question. \
+Specifically pay attention also to the entities, relationships and terms extracted, as these indicate what type of \
+objects/relationships/terms you can ask about! Do not ask about entities, terms or relationships that are not mentioned \
+in the 'entities, relationships and terms' section.
+
+Again, please find questions that are NOT overlapping too much with the already answered sub-questions or those that \
+already were suggested and failed. In other words - what can we try in addition to what has been tried so far?
+
+Generate the list of questions separated by one new line like this (and please ONLY ANSWER WITH THIS LIST! Do not \
+add any explanations or other text!):
+
 <sub-question 1>
 <sub-question 2>
 <sub-question 3>
@@ -489,7 +841,7 @@ Your task is to improve on a given answer to a question, as the initial answer w
 Use the information provided below - and only the provided information - to write your new and improved answer.
 
 The information provided below consists of:
-  1) an initial answer that was given but found to be lacking in some way.
+  1) an initial answer that was given but likely found to be lacking in some way.
   2) a number of answered sub-questions - these are very important(!) and definitely should help you to answer the main \
 question. Note that the sub-questions have a type, 'initial' and 'refined'. The 'initial' ones were available for the \
 creation of the initial answer, but the 'refined' were not, they are new. So please use the 'refined' sub-questions in \
@@ -499,6 +851,7 @@ particular to update/extend/correct/enrich the initial answer and to add more de
 the relevant document for a fact!
 
 It is critical that you provide proper inline citations to documents in the format [D1], [D2], [D3], etc! \
+Please use format [D1][D2] and NOT [D1, D2] format if you cite two or more documents together! \
 It is important that the citation is close to the information it supports. \
 DO NOT just list all of the citations at the very end. \
 Feel free to also cite sub-questions in addition to documents, \
@@ -509,14 +862,7 @@ Proper citations are very important for the user!
 
 {{history}}
 
-IMPORTANT RULES:
- - If you cannot reliably answer the question solely using the provided information, say that you cannot reliably answer. \
-You may give some additional facts you learned, but do not try to invent an answer.
- - If the information is empty or irrelevant, just say "{UNKNOWN_ANSWER}".
- - If the information is relevant but not fully conclusive, provide an answer to the extent you can but also specify that \
-the information is not conclusive and why.
- - Ignore any existing citations within the answered sub-questions, like [D1]... and [Q2]! The citations you will need to \
-use will need to refer to the documents (and sub-questions) that you are explicitly presented with below!
+{COMMON_RAG_RULES}
 
 Again, you should be sure that the answer is supported by the information provided!
 
@@ -545,7 +891,9 @@ Lastly, here is the main question I want you to answer based on the information
 {{question}}
 {SEPARATOR_LINE}
 
-Please keep your answer brief and concise, and focus on facts and data.
+Please keep your answer brief and concise, and focus on facts and data. (Again, only state what you see in the documents for \
+sure and communicate if/in what way this may or may not relate to the question you need to answer! Use the answered \
+sub-questions as well, but be cautious and reconsider the docments again for validation.)
 
 Answer:
 """.strip()
@@ -561,18 +909,13 @@ The information provided below consists of:
   2) a number of documents that were also deemed relevant for the question.
 
 It is critical that you provide proper inline citations to documents in the format [D1], [D2], [D3], etc! \
+Please use format [D1][D2] and NOT [D1, D2] format if you cite two or more documents together! \
 It is important that the citation is close to the information it supports. \
 DO NOT just list all of the citations at the very end of your response. Citations are very important for the user!
 
 {{history}}
 
-IMPORTANT RULES:
- - If you cannot reliably answer the question solely using the provided information, say that you cannot reliably answer. \
-You may give some additional facts you learned, but do not try to invent an answer.
- - If the information is empty or irrelevant, just say "{UNKNOWN_ANSWER}".
- - If the information is relevant but not fully conclusive, provide an answer to the extent you can but also specify that \
-the information is not conclusive and why.
-
+{COMMON_RAG_RULES}
 Again, you should be sure that the answer is supported by the information provided!
 
 Try to keep your answer concise. But also highlight uncertainties you may have should there be substantial ones, \
@@ -597,11 +940,103 @@ Lastly, here is the question I want you to answer based on the information above
 {{question}}
 {SEPARATOR_LINE}
 
-Please keep your answer brief and concise, and focus on facts and data.
+Please keep your answer brief and concise, and focus on facts and data. (Again, only state what you see in the documents for \
+sure and communicate if/in what way this may or may not relate to the question you need to answer!)
 
 Answer:
 """.strip()
 
+REFINED_ANSWER_VALIDATION_PROMPT = f"""
+{{persona_specification}}
+
+Your task is to verify whether a given answer is truthful and accurate, and supported by the facts that you \
+will be provided with.
+
+The information provided below consists of:
+
+  1) a question that needed to be answered
+
+  2) a proposed answer to the question, whose accuracy you should assess
+
+  3) potentially, a brief summary of the history of the conversation thus far, as it may give more context \
+to the question. Note that the statements in the history are NOT considered as facts, ONLY but serve to to \
+give context to the question.
+
+  4) a number of answered sub-questions - you can take the answers as facts for these purposes.
+
+  5) a number of relevant documents that should support the answer and that you should use as fact, \
+i.e., if a statement in the document backs up a statement in the answer, then that statement in the answer \
+should be considered as true.
+
+
+IMPORTANT RULES AND CONSIDERATIONS:
+
+ - Please consider the statements made in the proposed answer and assess whether they are truthful and accurate, based \
+on the provided sub-answered and the documents. (Again, the history is NOT considered as facts!)
+
+ - Look in particular for:
+    * material statements that are not supported by the sub-answered or the documents
+    * assignments and groupings that are not supported, like company A is competitor of company B, but this is not \
+explicitly supported by documents or sub-answers, guesses or interpretations unless explicitly asked for
+
+ - look also at the citations in the proposed answer and assess whether they are appropriate given the statements \
+made in the proposed answer that cites the document.
+
+ - Are items grouped together amongst one headline where not all items belong to the category of the headline? \
+(Example: "Products used by Company A" where some products listed are not used by Company A)
+
+ - Does the proposed answer address the question in full?
+
+ - Is the answer specific to the question? Example: if the question asks for the prices for products by Company A, \
+but the answer lists the prices for products by Company A and Company B, or products it cannot be sure are by \
+Company A, then this is not quite specific enough to the question and the answer should be rejected.
+
+- Similarly, if the question asks for properties of a certain class but the proposed answer lists or includes entities \
+that are not of that class without very explicitly saying so, then the answer should be considered inaccurate.
+
+ - If there are any calculations in the proposed answer that are not supported by the documents, they need to be tested. \
+If any calculation is wrong, the proposed answer should be considered as not trustworthy.
+
+
+Here is the information:
+{SEPARATOR_LINE_LONG}
+
+QUESTION:
+{SEPARATOR_LINE}
+{{question}}
+{SEPARATOR_LINE}
+
+PROPOSED ANSWER:
+{SEPARATOR_LINE}
+{{proposed_answer}}
+{SEPARATOR_LINE}
+
+Here is the additional contextual information:
+{SEPARATOR_LINE_LONG}
+
+{{history}}
+
+Sub-questions and their answers (to be considered as facts):
+{SEPARATOR_LINE}
+{{answered_sub_questions}}
+{SEPARATOR_LINE}
+
+And here are the relevant documents that support the sub-question answers, and that are relevant for the actual question:
+{SEPARATOR_LINE}
+{{relevant_docs}}
+{SEPARATOR_LINE}
+
+
+Please think through this step by step. Format your response just as a string in the following format:
+
+Analysis: <think through your reasoning as outlined in the 'IMPORTANT RULES AND CONSIDERATIONS' section above, \
+but keep it short. Come to a conclusion whether the proposed answer can be trusted>
+Comments: <state your condensed comments you would give to a user reading the proposed answer, regarding the accuracy and \
+specificity.>
+{AGENT_ANSWER_SEPARATOR} <answer here only with yes or no, whether the proposed answer can be trusted. Base this on your \
+analysis, but only say 'yes' (trustworthy) or 'no' (not trustworthy)>
+""".strip()
+
 
 INITIAL_REFINED_ANSWER_COMPARISON_PROMPT = f"""
 For the given question, please compare the initial answer and the refined answer and determine if the refined answer is \

backend/onyx/redis/redis_connector_credential_pair.py

@@ -120,6 +120,7 @@ class RedisConnectorCredentialPair(RedisObjectHelper):
                 queue=OnyxCeleryQueues.VESPA_METADATA_SYNC,
                 task_id=custom_task_id,
                 priority=OnyxCeleryPriority.MEDIUM,
+                ignore_result=True,
             )
 
             num_tasks_sent += 1

backend/onyx/redis/redis_connector_delete.py

@@ -132,6 +132,7 @@ class RedisConnectorDelete:
                 queue=OnyxCeleryQueues.CONNECTOR_DELETION,
                 task_id=custom_task_id,
                 priority=OnyxCeleryPriority.MEDIUM,
+                ignore_result=True,
             )
 
             async_results.append(result)

backend/onyx/redis/redis_connector_doc_perm_sync.py

@@ -11,6 +11,7 @@ from redis.lock import Lock as RedisLock
 
 from onyx.access.models import DocExternalAccess
 from onyx.configs.constants import CELERY_GENERIC_BEAT_LOCK_TIMEOUT
+from onyx.configs.constants import CELERY_PERMISSIONS_SYNC_LOCK_TIMEOUT
 from onyx.configs.constants import OnyxCeleryPriority
 from onyx.configs.constants import OnyxCeleryQueues
 from onyx.configs.constants import OnyxCeleryTask
@@ -49,7 +50,7 @@ class RedisConnectorPermissionSync:
     # it's impossible to get the exact state of the system at a single point in time
     # so we need a signal with a TTL to bridge gaps in our checks
     ACTIVE_PREFIX = PREFIX + "_active"
-    ACTIVE_TTL = 3600
+    ACTIVE_TTL = CELERY_PERMISSIONS_SYNC_LOCK_TIMEOUT * 2
 
     def __init__(self, tenant_id: str | None, id: int, redis: redis.Redis) -> None:
         self.tenant_id: str | None = tenant_id
@@ -195,6 +196,7 @@ class RedisConnectorPermissionSync:
                 queue=OnyxCeleryQueues.DOC_PERMISSIONS_UPSERT,
                 task_id=custom_task_id,
                 priority=OnyxCeleryPriority.HIGH,
+                ignore_result=True,
             )
             async_results.append(result)
 

backend/onyx/redis/redis_connector_prune.py

@@ -10,6 +10,7 @@ from redis.lock import Lock as RedisLock
 from sqlalchemy.orm import Session
 
 from onyx.configs.constants import CELERY_GENERIC_BEAT_LOCK_TIMEOUT
+from onyx.configs.constants import CELERY_PRUNING_LOCK_TIMEOUT
 from onyx.configs.constants import OnyxCeleryPriority
 from onyx.configs.constants import OnyxCeleryQueues
 from onyx.configs.constants import OnyxCeleryTask
@@ -49,7 +50,7 @@ class RedisConnectorPrune:
     # it's impossible to get the exact state of the system at a single point in time
     # so we need a signal with a TTL to bridge gaps in our checks
     ACTIVE_PREFIX = PREFIX + "_active"
-    ACTIVE_TTL = 3600
+    ACTIVE_TTL = CELERY_PRUNING_LOCK_TIMEOUT * 2
 
     def __init__(self, tenant_id: str | None, id: int, redis: redis.Redis) -> None:
         self.tenant_id: str | None = tenant_id
@@ -201,6 +202,7 @@ class RedisConnectorPrune:
                 queue=OnyxCeleryQueues.CONNECTOR_DELETION,
                 task_id=custom_task_id,
                 priority=OnyxCeleryPriority.MEDIUM,
+                ignore_result=True,
             )
 
             async_results.append(result)

backend/onyx/redis/redis_utils.py

@@ -0,0 +1,29 @@
+from onyx.redis.redis_connector_credential_pair import (
+    RedisGlobalConnectorCredentialPair,
+)
+from onyx.redis.redis_connector_delete import RedisConnectorDelete
+from onyx.redis.redis_connector_doc_perm_sync import RedisConnectorPermissionSync
+from onyx.redis.redis_connector_index import RedisConnectorIndex
+from onyx.redis.redis_connector_prune import RedisConnectorPrune
+from onyx.redis.redis_document_set import RedisDocumentSet
+from onyx.redis.redis_usergroup import RedisUserGroup
+
+
+def is_fence(key_bytes: bytes) -> bool:
+    key_str = key_bytes.decode("utf-8")
+    if key_str == RedisGlobalConnectorCredentialPair.FENCE_KEY:
+        return True
+    if key_str.startswith(RedisDocumentSet.FENCE_PREFIX):
+        return True
+    if key_str.startswith(RedisUserGroup.FENCE_PREFIX):
+        return True
+    if key_str.startswith(RedisConnectorDelete.FENCE_PREFIX):
+        return True
+    if key_str.startswith(RedisConnectorPrune.FENCE_PREFIX):
+        return True
+    if key_str.startswith(RedisConnectorIndex.FENCE_PREFIX):
+        return True
+    if key_str.startswith(RedisConnectorPermissionSync.FENCE_PREFIX):
+        return True
+
+    return False

backend/onyx/seeding/load_yamls.py

@@ -162,6 +162,11 @@ def load_personas_from_yaml(
                 else persona.get("is_visible")
             ),
             db_session=db_session,
+            is_default_persona=(
+                existing_persona.is_default_persona
+                if existing_persona is not None
+                else persona.get("is_default_persona", False)
+            ),
         )
 
 

backend/onyx/seeding/personas.yaml

@@ -41,6 +41,7 @@ personas:
     icon_color: "#6FB1FF"
     display_priority: 0
     is_visible: true
+    is_default_persona: true
     starter_messages:
       - name: "Give me an overview of what's here"
         message: "Sample some documents and tell me what you find."
@@ -66,6 +67,7 @@ personas:
     icon_color: "#FF6F6F"
     display_priority: 1
     is_visible: true
+    is_default_persona: true
     starter_messages:
       - name: "Summarize a document"
         message: "If I have provided a document please summarize it for me. If not, please ask me to upload a document either by dragging it into the input bar or clicking the +file icon."
@@ -91,6 +93,7 @@ personas:
     icon_color: "#6FFF8D"
     display_priority: 2
     is_visible: false
+    is_default_persona: true
     starter_messages:
       - name: "Document Search"
         message: "Hi! Could you help me find information about our team structure and reporting lines from our internal documents?"
@@ -117,6 +120,7 @@ personas:
     image_generation: true
     display_priority: 3
     is_visible: true
+    is_default_persona: true
     starter_messages:
       - name: "Create visuals for a presentation"
         message: "Generate someone presenting a graph which clearly demonstrates an upwards trajectory."

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

@@ -32,6 +32,7 @@ from onyx.db.persona import get_personas_for_user
 from onyx.db.persona import mark_persona_as_deleted
 from onyx.db.persona import mark_persona_as_not_deleted
 from onyx.db.persona import update_all_personas_display_priority
+from onyx.db.persona import update_persona_is_default
 from onyx.db.persona import update_persona_label
 from onyx.db.persona import update_persona_public_status
 from onyx.db.persona import update_persona_shared_users
@@ -56,7 +57,6 @@ from onyx.tools.utils import is_image_generation_available
 from onyx.utils.logger import setup_logger
 from onyx.utils.telemetry import create_milestone_and_report
 
-
 logger = setup_logger()
 
 
@@ -72,6 +72,10 @@ class IsPublicRequest(BaseModel):
     is_public: bool
 
 
+class IsDefaultRequest(BaseModel):
+    is_default_persona: bool
+
+
 @admin_router.patch("/{persona_id}/visible")
 def patch_persona_visibility(
     persona_id: int,
@@ -106,6 +110,25 @@ def patch_user_presona_public_status(
         raise HTTPException(status_code=403, detail=str(e))
 
 
+@admin_router.patch("/{persona_id}/default")
+def patch_persona_default_status(
+    persona_id: int,
+    is_default_request: IsDefaultRequest,
+    user: User | None = Depends(current_curator_or_admin_user),
+    db_session: Session = Depends(get_session),
+) -> None:
+    try:
+        update_persona_is_default(
+            persona_id=persona_id,
+            is_default=is_default_request.is_default_persona,
+            db_session=db_session,
+            user=user,
+        )
+    except ValueError as e:
+        logger.exception("Failed to update persona default status")
+        raise HTTPException(status_code=403, detail=str(e))
+
+
 @admin_router.put("/display-priority")
 def patch_persona_display_priority(
     display_priority_request: DisplayPriorityRequest,

backend/onyx/server/gpts/api.py

@@ -76,6 +76,7 @@ def gpt_search(
         user=None,
         llm=llm,
         fast_llm=fast_llm,
+        skip_query_analysis=True,
         db_session=db_session,
     ).reranked_sections
 

backend/onyx/server/settings/models.py

@@ -12,10 +12,10 @@ class PageType(str, Enum):
     SEARCH = "search"
 
 
-class GatingType(str, Enum):
-    FULL = "full"  # Complete restriction of access to the product or service
-    PARTIAL = "partial"  # Full access but warning (no credit card on file)
-    NONE = "none"  # No restrictions, full access to all features
+class ApplicationStatus(str, Enum):
+    PAYMENT_REMINDER = "payment_reminder"
+    GATED_ACCESS = "gated_access"
+    ACTIVE = "active"
 
 
 class Notification(BaseModel):
@@ -43,7 +43,7 @@ class Settings(BaseModel):
 
     maximum_chat_retention_days: int | None = None
     gpu_enabled: bool | None = None
-    product_gating: GatingType = GatingType.NONE
+    application_status: ApplicationStatus = ApplicationStatus.ACTIVE
     anonymous_user_enabled: bool | None = None
     pro_search_disabled: bool | None = None
     auto_scroll: bool | None = None