fix(vector-stores): fall back to annotate when the failure mode is unrecognized

litellm_settings keys are set on the litellm module with no allowlist, so a typo
in vector_store_search_failure_mode reached assert_never and turned every
vector-store request into a 500. Validate the configured value and fall back to
the permissive default with a warning naming the supported modes.

litellm/integrations/vector_store_integrations/vector_store_pre_call_hook.py

@@ -7,8 +7,9 @@ It searches the vector store for relevant context and appends it to the messages
 
 from collections.abc import Awaitable, Callable, Mapping, Sequence
 from dataclasses import dataclass
-from typing import TYPE_CHECKING, Any, Final, Protocol, cast
+from typing import TYPE_CHECKING, Any, Final, Protocol, cast, get_args
 
+from pydantic import TypeAdapter, ValidationError
 from typing_extensions import assert_never
 
 import litellm
@@ -22,6 +23,7 @@ from litellm.types.utils import CallTypes, StandardCallbackDynamicParams
 from litellm.types.vector_stores import (
     LiteLLM_ManagedVectorStore,
     VectorStoreSearchFailure,
+    VectorStoreSearchFailureMode,
     VectorStoreSearchResponse,
     VectorStoreSearchResult,
 )
@@ -34,6 +36,8 @@ else:
     LiteLLMLoggingObj = Any
 
 SEARCH_FAILURES_FIELD: Final = "vector_store_search_failures"
+_DEFAULT_FAILURE_MODE: Final[VectorStoreSearchFailureMode] = "annotate"
+_FAILURE_MODE_ADAPTER: Final = TypeAdapter(VectorStoreSearchFailureMode)
 
 
 class ProxyRuntime(Protocol):
@@ -153,7 +157,7 @@ class VectorStorePreCallHook(CustomLogger):
                 litellm_logging_obj.model_call_details[detail] = value
 
         if augmentation.failures:
-            match litellm.vector_store_search_failure_mode:
+            match _configured_failure_mode():
                 case "error":
                     raise VectorStoreSearchError(failures=augmentation.failures, model=model)
                 case "annotate":
@@ -435,3 +439,16 @@ def _requested_vector_store_ids(non_default_params: Mapping[str, object]) -> tup
     if not isinstance(requested, (list, tuple)):
         return ()
     return tuple(str(vector_store_id) for vector_store_id in requested)
+
+
+def _configured_failure_mode() -> VectorStoreSearchFailureMode:
+    try:
+        return _FAILURE_MODE_ADAPTER.validate_python(litellm.vector_store_search_failure_mode)
+    except ValidationError:
+        verbose_logger.warning(
+            "Unsupported vector_store_search_failure_mode=%r, falling back to %r. Supported modes: %s",
+            litellm.vector_store_search_failure_mode,
+            _DEFAULT_FAILURE_MODE,
+            ", ".join(get_args(VectorStoreSearchFailureMode)),
+        )
+        return _DEFAULT_FAILURE_MODE

tests/test_litellm/integrations/vector_store_integrations/test_vector_store_pre_call_hook.py

@@ -425,6 +425,36 @@ async def test_error_mode_fails_the_request_instead_of_answering_without_the_kno
     assert "vs-broken: litellm.BadRequestError: no healthy deployments for vs-broken" in raised.value.message
 
 
+@pytest.mark.asyncio
+async def test_a_misspelled_failure_mode_annotates_instead_of_erroring_the_request(
+    registry_with: RegisterStores,
+    monkeypatch: pytest.MonkeyPatch,
+    warnings: list[logging.LogRecord],
+) -> None:
+    """Regression (LIT-6809): litellm_settings takes any value, so a typo must not become a 500."""
+    registry_with("vs-broken")
+    monkeypatch.setattr(litellm, "vector_store_search_failure_mode", "erorr")
+
+    logging_obj = FakeLoggingObj({})
+    _, messages, _ = await _run_hook(
+        VectorStorePreCallHook(
+            proxy_runtime=FakeProxyRuntime(router=RecordingRouter(failing_vector_store_ids=frozenset({"vs-broken"})))
+        ),
+        ["vs-broken"],
+        logging_obj,
+    )
+
+    assert messages[0]["content"] == "what is litellm?"
+    assert logging_obj.model_call_details["vector_store_search_failures"] == (
+        {
+            "vector_store_id": "vs-broken",
+            "custom_llm_provider": "bedrock",
+            "error": "litellm.BadRequestError: no healthy deployments for vs-broken",
+        },
+    )
+    assert any("erorr" in record.getMessage() for record in warnings)
+
+
 @pytest.mark.asyncio
 async def test_error_mode_leaves_a_fully_healthy_request_alone(
     registry_with: RegisterStores,