Merge pull request #33649 from BerriAI/litellm_list_vs_fil

fix(proxy): resolve team wildcard credentials for vector store files

litellm/proxy/vector_store_files_endpoints/endpoints.py

@@ -227,6 +227,8 @@ async def _update_request_data_with_model_routing_hint(
     model_hint = data.get("model") or user_controlled_model_hint
     should_authorize_model_hint = isinstance(model_hint, str) and model_hint == user_controlled_model_hint
 
+    caller_team_id = getattr(user_api_key_dict, "team_id", None) if user_api_key_dict else None
+
     should_route = False
     credentials = None
     if isinstance(model_hint, str) and "*" in model_hint:
@@ -237,7 +239,9 @@ async def _update_request_data_with_model_routing_hint(
                     llm_router=llm_router,
                     user_api_key_dict=user_api_key_dict,
                 )
-            credentials = llm_router.get_deployment_credentials_with_provider(model_id=model_hint)
+            credentials = llm_router.get_deployment_credentials_with_provider(
+                model_id=model_hint, team_id=caller_team_id
+            )
             should_route = credentials is not None
     else:
         if isinstance(model_hint, str) and should_authorize_model_hint:
@@ -285,7 +289,7 @@ async def _update_request_data_with_model_routing_hint(
 
     openai_credentials = None
     for model_name in model_names_to_check:
-        credentials = llm_router.get_deployment_credentials_with_provider(model_id=model_name)
+        credentials = llm_router.get_deployment_credentials_with_provider(model_id=model_name, team_id=caller_team_id)
         if credentials is None:
             continue
 

litellm/router.py

@@ -7940,6 +7940,7 @@ class Router:
         self.model_id_to_deployment_index_map = {}  # Reset the index
         self.model_name_to_deployment_indices = {}  # Reset the model_name index
         self.team_model_to_deployment_indices = {}  # Reset the team_model index
+        self.team_pattern_routers = {}
         self.team_public_model_names = frozenset()
         # Reset per-strategy router registries so hot-reload doesn't leave
         # stale routers pointing at the old model_list.
@@ -8296,6 +8297,12 @@ class Router:
             public_model_name for _, public_model_name in self.team_model_to_deployment_indices
         )
 
+        for team_id in list(self.team_pattern_routers.keys()):
+            team_pattern_router = self.team_pattern_routers[team_id]
+            team_pattern_router.remove_deployment(model_id)
+            if not team_pattern_router.patterns:
+                del self.team_pattern_routers[team_id]
+
     def _update_team_model_index(self, model: dict, idx: int) -> None:
         """
         Helper to update team_model_to_deployment_indices for a single deployment.
@@ -8544,7 +8551,9 @@ class Router:
             int(max_output) if max_output is not None else None,
         )
 
-    def get_deployment_credentials_with_provider(self, model_id: str) -> Optional[Dict[str, Any]]:
+    def get_deployment_credentials_with_provider(
+        self, model_id: str, team_id: str | None = None
+    ) -> dict[str, Any] | None:
         """
         Get API credentials and provider info from a model name in model_list.
         Useful for passthrough endpoints (files, batches, etc.) that need credentials.
@@ -8554,6 +8563,9 @@ class Router:
 
         Args:
             model_id: Model ID or model name from model_list (e.g., "gpt-4o-litellm")
+            team_id: Optional team id of the caller. When set, team-scoped
+                deployments (indexed by team public model name, including team
+                wildcard models like "openai/*") are also considered.
 
         Returns:
             Dictionary containing api_key, api_base, custom_llm_provider, etc.
@@ -8572,9 +8584,22 @@ class Router:
         if deployment is None:
             deployment = self.get_deployment_by_model_group_name(model_group_name=model_id)
 
-        # If still not found, check for wildcard pattern matches
+        # If not found, check team-scoped deployments whose team public model
+        # name exactly matches model_id (wildcard team names are matched via
+        # team_pattern_routers below).
+        if deployment is None and team_id is not None:
+            team_indices = self.team_model_to_deployment_indices.get((team_id, model_id), [])
+            if team_indices:
+                team_model = self.model_list[team_indices[0]]
+                deployment = Deployment(**team_model) if isinstance(team_model, dict) else team_model
+
+        # If still not found, check for wildcard pattern matches. Team wildcard
+        # matches take priority so a global pattern (e.g. "openai/*") doesn't
+        # shadow the team's own entry.
         if deployment is None:
-            potential_wildcard_models = self.pattern_router.route(model_id) or []
+            team_pattern_router = self.team_pattern_routers.get(team_id) if team_id is not None else None
+            team_wildcard_models = (team_pattern_router.route(model_id) or []) if team_pattern_router else []
+            potential_wildcard_models = team_wildcard_models or self.pattern_router.route(model_id) or []
             if potential_wildcard_models:
                 # Use the first matching wildcard deployment
                 deployment_dict = potential_wildcard_models[0]

litellm/router_utils/pattern_match_deployments.py

@@ -73,6 +73,17 @@ class PatternMatchRouter:
             self.patterns[regex] = []
         self.patterns[regex].append(llm_deployment)
 
+    def remove_deployment(self, model_id: str) -> None:
+        """
+        Remove every deployment with the given model id from the pattern registry,
+        dropping any pattern whose deployment list becomes empty.
+        """
+        self.patterns = {
+            regex: remaining
+            for regex, deployments in self.patterns.items()
+            if (remaining := [d for d in deployments if (d.get("model_info") or {}).get("id") != model_id])
+        }
+
     def _pattern_to_regex(self, pattern: str) -> str:
         """
         Convert a wildcard pattern to a regex pattern

tests/test_litellm/proxy/vector_store_endpoints/test_vector_store_endpoints.py

@@ -210,7 +210,7 @@ async def test_vector_store_file_list_resolves_single_openai_team_deployment():
     assert result["model"] == "openai/gpt-4o-mini"
     assert "custom_llm_provider" not in result
     llm_router.get_deployment_credentials_with_provider.assert_called_once_with(
-        model_id="team-openai"
+        model_id="team-openai", team_id=None
     )
 
 

tests/test_litellm/test_router.py

@@ -3714,6 +3714,125 @@ def test_get_deployment_credentials_with_provider_resolves_credential_name():
     litellm.credential_list = []
 
 
+def _team_wildcard_model(api_key: str, model_id: str = "team-wildcard-id") -> dict:
+    return {
+        "model_name": f"model_name_team-1_{model_id}",
+        "litellm_params": {"model": "openai/*", "api_key": api_key},
+        "model_info": {
+            "id": model_id,
+            "team_id": "team-1",
+            "team_public_model_name": "openai/*",
+        },
+    }
+
+
+def test_get_deployment_credentials_with_provider_team_wildcard_priority():
+    """
+    Regression: a global wildcard pattern (e.g. "openai/*") must not shadow a
+    team's own wildcard entry. When team_id is provided, the team wildcard
+    deployment's credentials win; without team_id the global one is used.
+    """
+    router = litellm.Router(
+        model_list=[
+            {
+                "model_name": "openai/*",
+                "litellm_params": {"model": "openai/*", "api_key": "global-key"},
+            },
+            _team_wildcard_model(api_key="team-key"),
+        ],
+    )
+
+    team_credentials = router.get_deployment_credentials_with_provider(
+        model_id="openai/gpt-5.2", team_id="team-1"
+    )
+    assert team_credentials is not None
+    assert team_credentials["api_key"] == "team-key"
+
+    global_credentials = router.get_deployment_credentials_with_provider(
+        model_id="openai/gpt-5.2"
+    )
+    assert global_credentials is not None
+    assert global_credentials["api_key"] == "global-key"
+
+
+def test_team_wildcard_credentials_not_usable_after_delete_deployment():
+    """
+    Regression: team_pattern_routers retained deleted deployments, so a team
+    user could keep resolving credentials of a deleted wildcard deployment.
+    """
+    router = litellm.Router(model_list=[_team_wildcard_model(api_key="old-key")])
+
+    assert (
+        router.get_deployment_credentials_with_provider(
+            model_id="openai/gpt-5.2", team_id="team-1"
+        )
+        is not None
+    )
+
+    router.delete_deployment(id="team-wildcard-id")
+
+    assert (
+        router.get_deployment_credentials_with_provider(
+            model_id="openai/gpt-5.2", team_id="team-1"
+        )
+        is None
+    )
+
+
+def test_pattern_match_router_remove_deployment():
+    """
+    remove_deployment must drop only the deployment with the given model id and
+    delete patterns whose deployment list becomes empty.
+    """
+    from litellm.router_utils.pattern_match_deployments import PatternMatchRouter
+
+    pattern_router = PatternMatchRouter()
+    pattern_router.add_pattern(
+        "openai/*",
+        {"litellm_params": {"model": "openai/*", "api_key": "key-a"}, "model_info": {"id": "dep-a"}},
+    )
+    pattern_router.add_pattern(
+        "openai/*",
+        {"litellm_params": {"model": "openai/*", "api_key": "key-b"}, "model_info": {"id": "dep-b"}},
+    )
+
+    pattern_router.remove_deployment(model_id="dep-a")
+    matches = pattern_router.route("openai/gpt-5.2")
+    assert matches is not None
+    assert [m["model_info"]["id"] for m in matches] == ["dep-b"]
+
+    pattern_router.remove_deployment(model_id="dep-b")
+    assert pattern_router.patterns == {}
+    assert pattern_router.route("openai/gpt-5.2") is None
+
+
+def test_team_wildcard_credentials_refreshed_on_upsert_and_set_model_list():
+    """
+    Regression: replacing a team wildcard deployment (upsert or model list
+    reload) must serve the new credentials, not the stale cached ones.
+    """
+    from litellm.types.router import Deployment
+
+    router = litellm.Router(model_list=[_team_wildcard_model(api_key="old-key")])
+
+    router.upsert_deployment(
+        deployment=Deployment(**_team_wildcard_model(api_key="new-key"))
+    )
+    credentials = router.get_deployment_credentials_with_provider(
+        model_id="openai/gpt-5.2", team_id="team-1"
+    )
+    assert credentials is not None
+    assert credentials["api_key"] == "new-key"
+
+    router.set_model_list(model_list=[])
+    assert (
+        router.get_deployment_credentials_with_provider(
+            model_id="openai/gpt-5.2", team_id="team-1"
+        )
+        is None
+    )
+
+
 def test_get_available_guardrail_single_deployment():
     """
     Test get_available_guardrail returns the single guardrail when only one exists.