fix(vector-stores): approve config gRPC stores with prefixed providers

approve_configured_connection only stamped the admin marker when the
provider was exactly "milvus", while gRPC detection and the managed
rejection both normalize prefixed values such as "milvus/team", so a
config-defined gRPC store with a prefixed provider was rejected with 403
on every search. Normalize the provider the same way before stamping

Also give VectorStoreRegistry and VectorStoreIndexRegistry their own
list per instance instead of a shared mutable default, which made every
registry built without an explicit list see the stores loaded into any
other one. The budget files are ratcheted from the merge-base values, so
the earlier merge commit's double-lowered limits are corrected here

basedpyright-code-budget.json

@@ -1,9 +1,9 @@
 {
   "reportAny": {
-    "limit": 13422
+    "limit": 13427
   },
   "reportArgumentType": {
-    "limit": 2137
+    "limit": 2187
   },
   "reportAssignmentType": {
     "limit": 319
@@ -24,7 +24,7 @@
     "limit": 19
   },
   "reportExplicitAny": {
-    "limit": 3363
+    "limit": 3368
   },
   "reportFunctionMemberAccess": {
     "limit": 7
@@ -57,7 +57,7 @@
     "limit": 5570
   },
   "reportMissingTypeArgument": {
-    "limit": 15246
+    "limit": 15275
   },
   "reportMissingTypeStubs": {
     "limit": 40
@@ -99,19 +99,19 @@
     "limit": 0
   },
   "reportUnknownArgumentType": {
-    "limit": 42551
+    "limit": 44133
   },
   "reportUnknownLambdaType": {
     "limit": 109
   },
   "reportUnknownMemberType": {
-    "limit": 38191
+    "limit": 38267
   },
   "reportUnknownParameterType": {
-    "limit": 19576
+    "limit": 19582
   },
   "reportUnknownVariableType": {
-    "limit": 29814
+    "limit": 29826
   },
   "reportUnnecessaryCast": {
     "limit": 110
@@ -123,7 +123,7 @@
     "limit": 4
   },
   "reportUnnecessaryIsInstance": {
-    "limit": 810
+    "limit": 815
   },
   "reportUntypedBaseClass": {
     "limit": 0

litellm/llms/milvus/vector_stores/connection.py

@@ -125,6 +125,6 @@ def managed_connection_fields(custom_llm_provider: object) -> frozenset[str]:
 def approve_configured_connection(
     custom_llm_provider: object, litellm_params: Mapping[str, object]
 ) -> Mapping[str, object]:
-    if custom_llm_provider == "milvus" and litellm_params.get("milvus_transport") == "grpc":
+    if _normalize_provider(custom_llm_provider) == "milvus" and litellm_params.get("milvus_transport") == "grpc":
         return MappingProxyType({**litellm_params, MILVUS_ADMIN_CONFIGURED_CONNECTION: True})
     return litellm_params

litellm/vector_stores/vector_store_registry.py

@@ -58,8 +58,10 @@ def _normalized_vector_store(vector_store: LiteLLM_ManagedVectorStore) -> LiteLL
 
 
 class VectorStoreIndexRegistry:
-    def __init__(self, vector_store_indexes: list[LiteLLM_ManagedVectorStoreIndex] = []):
-        self.vector_store_indexes: list[LiteLLM_ManagedVectorStoreIndex] = vector_store_indexes
+    def __init__(self, vector_store_indexes: list[LiteLLM_ManagedVectorStoreIndex] | None = None):
+        self.vector_store_indexes: list[LiteLLM_ManagedVectorStoreIndex] = (
+            vector_store_indexes if vector_store_indexes is not None else []
+        )
 
     def get_vector_store_indexes(self) -> list[LiteLLM_ManagedVectorStoreIndex]:
         """
@@ -127,8 +129,8 @@ class VectorStoreIndexRegistry:
 
 
 class VectorStoreRegistry:
-    def __init__(self, vector_stores: list[LiteLLM_ManagedVectorStore] = []):
-        self.vector_stores: list[LiteLLM_ManagedVectorStore] = vector_stores
+    def __init__(self, vector_stores: list[LiteLLM_ManagedVectorStore] | None = None):
+        self.vector_stores: list[LiteLLM_ManagedVectorStore] = vector_stores if vector_stores is not None else []
         self.vector_store_ids_to_vector_store_map: dict[str, LiteLLM_ManagedVectorStore] = {}
         self.config_vector_store_ids: frozenset[str] = frozenset()
 

ruff-strict-budget.json

@@ -33,7 +33,7 @@
     "limit": 2
   },
   "B006": {
-    "limit": 176
+    "limit": 174
   },
   "B008": {
     "limit": 503
@@ -246,7 +246,7 @@
     "limit": 109
   },
   "TRY300": {
-    "limit": 848
+    "limit": 851
   },
   "UP028": {
     "limit": 2

test-quality-budget.json

@@ -21,6 +21,6 @@
     "limit": 117
   },
   "TQ008": {
-    "limit": 10972
+    "limit": 10990
   }
 }

tests/test_litellm/proxy/vector_store_endpoints/test_vector_store_endpoints.py

@@ -794,13 +794,14 @@ async def test_unmarked_managed_milvus_connection_requires_admin_resave():
 
 
 @pytest.mark.asyncio
-async def test_config_loaded_milvus_grpc_connection_is_trusted():
+@pytest.mark.parametrize("custom_llm_provider", ["milvus", "milvus/probe"])
+async def test_config_loaded_milvus_grpc_connection_is_trusted(custom_llm_provider: str):
     registry = VectorStoreRegistry()
     source = {
         "vector_store_name": "configured",
         "litellm_params": {
             "vector_store_id": "configured",
-            "custom_llm_provider": "milvus",
+            "custom_llm_provider": custom_llm_provider,
             "milvus_transport": "grpc",
             "api_base": "https://configured-milvus:19530",
             "litellm_embedding_model": "team-embedding-alias",

tests/test_litellm/vector_stores/test_vector_store_registry.py

@@ -105,6 +105,19 @@ def test_get_credentials_for_vector_store():
     assert result == {}
 
 
+def test_fresh_registries_do_not_share_config_loaded_stores():
+    VectorStoreRegistry().load_vector_stores_from_config(
+        [
+            {
+                "vector_store_name": "configured",
+                "litellm_params": {"vector_store_id": "configured", "custom_llm_provider": "openai"},
+            }
+        ]
+    )
+
+    assert VectorStoreRegistry().get_litellm_managed_vector_store_from_registry("configured") is None
+
+
 def test_add_vector_store_to_registry():
     """Test that add_vector_store_to_registry adds vector store correctly when there are pre-existing stores"""
     # Create pre-existing vector stores