fix(proxy): harden Vertex pass-through route matching against substring false positives (#45787)

A plain pass-through endpoint whose target URL contains the substring
"predict", "search", "embedContent", "rawPredict", "streamGenerateContent"
or "predictLongRunning" could be misrouted to the Vertex AI logging handler.
Inside vertex_passthrough_handler the unguarded response["predictions"]
access then raised KeyError in the logging worker and silently dropped the
spend row (issue #45787).

- vertex_passthrough_logging_handler.py: match Vertex method routes by the
  canonical ":method" suffix (mirroring PassThroughEndpointLogging.is_vertex_route)
  instead of bare substring matching; the batchPredictionJobs resource route
  is path-based and kept as-is.
- vertex_embeddings/transformation.py: raise a clear ValueError when the
  embedding response lacks the "predictions" field instead of a raw KeyError
  surfacing deep in the logging worker.
- test_success_handler.py: regression tests for the issue URLs (negative) and
  for authentic Vertex ":method" URLs (positive).

AI-assisted disclosure: root cause located and patch written with the help of an
AI assistant; all tests and lint verified locally.

Signed-off-by: apex-mochen <2756823972@qq.com>

litellm/llms/vertex_ai/vertex_embeddings/transformation.py

@@ -229,6 +229,9 @@ class VertexAITextEmbeddingConfig(LiteLLMBaseModel):
         if VertexBGEConfig.is_bge_model(model):
             return VertexBGEConfig.transform_response(response=response, model=model, model_response=model_response)
 
+        if not isinstance(response, dict) or "predictions" not in response:
+            raise ValueError("Vertex embedding response missing 'predictions' field")
+
         _predictions: Final = response["predictions"]
 
         embedding_response: Final = []
@@ -257,6 +260,9 @@ class VertexAITextEmbeddingConfig(LiteLLMBaseModel):
         """
         Transforms a vertex fine-tuned model embedding response to an openai response format.
         """
+        if not isinstance(response, dict) or "predictions" not in response:
+            raise ValueError("Vertex embedding response missing 'predictions' field")
+
         _predictions: Final = response["predictions"]
 
         embedding_response: Final = []

litellm/proxy/pass_through_endpoints/llm_provider_handlers/vertex_passthrough_logging_handler.py

@@ -148,7 +148,7 @@ class VertexPassthroughLoggingHandler:
                 custom_llm_provider="vertex_ai",
                 vertex_location=vertex_location,
             )
-        if "predictLongRunning" in url_route:
+        if ":predictLongRunning" in url_route:
             model = VertexPassthroughLoggingHandler.extract_model_from_url(url_route)
 
             vertex_video_config: Final = VertexAIVideoConfig()
@@ -188,7 +188,7 @@ class VertexPassthroughLoggingHandler:
                 "kwargs": kwargs,
             }
 
-        elif "generateContent" in url_route:
+        elif ":generateContent" in url_route or ":streamGenerateContent" in url_route:
             model = VertexPassthroughLoggingHandler.extract_model_from_url(url_route)
 
             instance_of_vertex_llm: Final = litellm.VertexGeminiConfig()
@@ -220,7 +220,7 @@ class VertexPassthroughLoggingHandler:
                 "kwargs": kwargs,
             }
 
-        elif "embedContent" in url_route or "batchEmbedContents" in url_route:
+        elif ":embedContent" in url_route or ":batchEmbedContents" in url_route:
             return VertexPassthroughLoggingHandler._handle_embed_content_response(
                 httpx_response=httpx_response,
                 logging_obj=logging_obj,
@@ -228,14 +228,14 @@ class VertexPassthroughLoggingHandler:
                 kwargs=kwargs,
                 request_body=request_body,
             )
-        elif "predict" in url_route:
+        elif ":predict" in url_route:
             return VertexPassthroughLoggingHandler._handle_predict_response(
                 httpx_response=httpx_response,
                 logging_obj=logging_obj,
                 url_route=url_route,
                 kwargs=kwargs,
             )
-        elif "rawPredict" in url_route or "streamRawPredict" in url_route:
+        elif ":rawPredict" in url_route or ":streamRawPredict" in url_route:
             from litellm.llms.vertex_ai.vertex_ai_partner_models import (
                 get_vertex_ai_partner_model_config,
             )
@@ -287,7 +287,7 @@ class VertexPassthroughLoggingHandler:
                 "result": litellm_prediction_response,
                 "kwargs": kwargs,
             }
-        elif "search" in url_route:
+        elif ":search" in url_route:
             litellm_vs_response: Final = vertex_search_api_config.transform_search_vector_store_response(
                 response=httpx_response,
                 litellm_logging_obj=logging_obj,
@@ -513,7 +513,7 @@ class VertexPassthroughLoggingHandler:
 
         model: Final = VertexPassthroughLoggingHandler.extract_model_from_url(url_route)
         response_json: Final = httpx_response.json()
-        is_batch: Final = "batchEmbedContents" in url_route
+        is_batch: Final = ":batchEmbedContents" in url_route
 
         input_text: Final = VertexPassthroughLoggingHandler._extract_embed_content_input(
             request_body=request_body, batch=is_batch
@@ -626,7 +626,7 @@ class VertexPassthroughLoggingHandler:
         url_route: str,
     ) -> ModelResponse | TextCompletionResponse | None:
         parsed_chunks = []
-        if "generateContent" in url_route or "streamGenerateContent" in url_route:
+        if ":generateContent" in url_route or ":streamGenerateContent" in url_route:
             vertex_iterator: Any = VertexModelResponseIterator(
                 streaming_response=None,
                 sync_stream=False,
@@ -634,7 +634,7 @@ class VertexPassthroughLoggingHandler:
             )
             chunk_parsing_logic: Callable[..., ModelResponseStream | None] = vertex_iterator._common_chunk_parsing_logic
             parsed_chunks = [chunk_parsing_logic(chunk) for chunk in all_chunks]
-        elif "rawPredict" in url_route or "streamRawPredict" in url_route:
+        elif ":rawPredict" in url_route or ":streamRawPredict" in url_route:
             from litellm.llms.anthropic.chat.handler import ModelResponseIterator
             from litellm.llms.base_llm.base_model_iterator import (
                 BaseModelResponseIterator,

tests/unit/proxy/pass_through_endpoints/test_success_handler.py

@@ -584,3 +584,68 @@ def test_vertex_ai_live_route_without_usage_frames_yields_no_logging_response(
 
     assert normalized["standard_logging_response_object"] is None
     assert normalized["kwargs"] == {"model": _LIVE_MODEL}
+
+
+# --- regression for #45787: substring route matching must not classify plain
+# passthrough URLs (e.g. "http://example.com/api/predict/laya") as Vertex AI ---
+
+_NON_VERTEX_URLS_WITH_METHOD_SUBSTRINGS: Final = [
+    "http://example.com/api/predict/laya",
+    "http://example.com/api/search/items",
+    "http://example.com/api/embedContent/proxy",
+    "http://example.com/v1/rawPredict",
+    "http://example.com/predictLongRunning/jobs/1",
+    "http://example.com/api/streamGenerateContent/relay",
+]
+
+
+@pytest.mark.parametrize("url", _NON_VERTEX_URLS_WITH_METHOD_SUBSTRINGS)
+def test_is_vertex_route_ignores_non_vertex_urls_containing_method_substrings(url: str) -> None:
+    assert PassThroughEndpointLogging().is_vertex_route(url) is False
+
+
+def test_is_vertex_route_matches_authentic_vertex_method_urls() -> None:
+    handler = PassThroughEndpointLogging()
+    assert (
+        handler.is_vertex_route(
+            "https://us-central1-aiplatform.googleapis.com/v1/projects/p/locations/us-central1/"
+            "publishers/google/models/gemini-2.0-flash:predict"
+        )
+        is True
+    )
+    assert (
+        handler.is_vertex_route(
+            "https://us-central1-aiplatform.googleapis.com/v1/projects/p/locations/us-central1/"
+            "publishers/google/models/gemini-2.0:search"
+        )
+        is True
+    )
+    assert (
+        handler.is_vertex_route(
+            "https://us-central1-aiplatform.googleapis.com/v1/projects/p/locations/us-central1/"
+            "publishers/google/models/text-embedding-005:embedContent"
+        )
+        is True
+    )
+
+
+@pytest.mark.usefixtures("local_model_cost_map")
+def test_non_vertex_predict_url_normalizes_without_predictions_keyerror() -> None:
+    # Issue #45787: a plain passthrough URL containing the substring "predict"
+    # used to be misrouted to the Vertex handler, whose unguarded
+    # response["predictions"] access then raised KeyError inside the logging
+    # worker and silently dropped the spend row.
+    normalized = PassThroughEndpointLogging().normalize_llm_passthrough_logging_payload(
+        httpx_response=httpx.Response(200, request=httpx.Request("GET", "http://example.com/api/predict/laya")),
+        response_body={},  # no "predictions" key
+        request_body={},
+        logging_obj=_logging_obj(),
+        url_route="http://example.com/api/predict/laya",
+        result="",
+        start_time=_START,
+        end_time=_END,
+        cache_hit=False,
+        custom_llm_provider=None,
+    )
+
+    assert normalized["standard_logging_response_object"] is not None