Merge pull request #37425 from BerriAI/litellm_fix_passthrough_embeddings_unmapped_spend

fix(proxy): log spend for OpenAI passthrough embeddings with unmapped models

litellm/proxy/pass_through_endpoints/llm_provider_handlers/openai_passthrough_logging_handler.py

@@ -229,6 +229,25 @@ class OpenAIPassthroughLoggingHandler(BasePassthroughLoggingHandler):
             verbose_proxy_logger.warning("Error calculating image editing cost: %s", e)
             return 0.0
 
+    @staticmethod
+    def _calculate_embeddings_cost(
+        litellm_model_response: EmbeddingResponse,
+        model: str,
+        custom_llm_provider: str,
+    ) -> float:
+        try:
+            return litellm.completion_cost(
+                completion_response=litellm_model_response,
+                model=model,
+                custom_llm_provider=custom_llm_provider,
+                call_type="aembedding",
+            )
+        except Exception as e:  # noqa: BLE001  # completion_cost raises bare Exception for unmapped models; cost failure must never drop the spend log
+            verbose_proxy_logger.warning(
+                "Error calculating embeddings cost for model %s, logging spend with cost 0: %s", model, e
+            )
+            return 0.0
+
     @staticmethod
     def _build_responses_api_response_and_cost(
         model: str,
@@ -351,11 +370,10 @@ class OpenAIPassthroughLoggingHandler(BasePassthroughLoggingHandler):
                     model_response_object=EmbeddingResponse(),
                     response_type="embedding",
                 )
-                response_cost = litellm.completion_cost(
-                    completion_response=litellm_model_response,
+                response_cost = OpenAIPassthroughLoggingHandler._calculate_embeddings_cost(
+                    litellm_model_response=litellm_model_response,
                     model=model,
                     custom_llm_provider=custom_llm_provider,
-                    call_type="aembedding",
                 )
                 litellm_model_response._hidden_params["response_cost"] = response_cost
             elif is_image_generation:
@@ -471,6 +489,12 @@ class OpenAIPassthroughLoggingHandler(BasePassthroughLoggingHandler):
 
         except Exception as e:
             verbose_proxy_logger.error("Error in OpenAI passthrough cost tracking: %s", e)
+            if not is_chat_completions:
+                unbilled_result: Final[PassThroughEndpointLoggingTypedDict] = {
+                    "result": None,
+                    "kwargs": kwargs,
+                }
+                return unbilled_result
             # Fall back to base handler without cost tracking
             base_handler = OpenAIPassthroughLoggingHandler()
             return base_handler.passthrough_chat_handler(

tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_openai_passthrough_logging_handler.py

@@ -1286,6 +1286,75 @@ class TestOpenAIPassthroughIntegration:
         mock_chat_handler.assert_called_once()
         assert result == {"result": None, "kwargs": {}}
 
+    def test_openai_passthrough_handler_embeddings_unmapped_model_logs_zero_cost(self):
+        response_body = {
+            "object": "list",
+            "model": "lit5787-unmapped-embeddings-deployment",
+            "data": [{"object": "embedding", "index": 0, "embedding": [0.1, 0.2]}],
+            "usage": {"prompt_tokens": 9, "total_tokens": 9},
+        }
+        mock_logging_obj = self._create_mock_logging_obj()
+        result = OpenAIPassthroughLoggingHandler.openai_passthrough_handler(
+            httpx_response=self._create_mock_httpx_response(response_body),
+            response_body=response_body,
+            logging_obj=mock_logging_obj,
+            url_route="https://my-resource.openai.azure.com/openai/v1/embeddings",
+            result="",
+            start_time=self.start_time,
+            end_time=self.end_time,
+            cache_hit=False,
+            request_body={
+                "model": "lit5787-unmapped-embeddings-deployment",
+                "input": "spend probe",
+            },
+            passthrough_logging_payload=PassthroughStandardLoggingPayload(
+                url="https://my-resource.openai.azure.com/openai/v1/embeddings",
+                request_body={
+                    "model": "lit5787-unmapped-embeddings-deployment",
+                    "input": "spend probe",
+                },
+                request_method="POST",
+            ),
+            litellm_params={},
+        )
+
+        assert result["result"] is not None
+        assert result["result"].usage.prompt_tokens == 9
+        assert result["kwargs"]["response_cost"] == 0.0
+        assert result["kwargs"]["model"] == "lit5787-unmapped-embeddings-deployment"
+        assert result["result"]._hidden_params["response_cost"] == 0.0
+        assert mock_logging_obj.model_call_details["response_cost"] == 0.0
+
+    def test_openai_passthrough_handler_embeddings_error_skips_chat_fallback(self):
+        response_body = {
+            "object": "list",
+            "model": "text-embedding-3-small",
+            "usage": {"prompt_tokens": 9, "total_tokens": 9},
+        }
+        kwargs_in = {
+            "passthrough_logging_payload": PassthroughStandardLoggingPayload(
+                url="https://api.openai.com/v1/embeddings",
+                request_body={"model": "text-embedding-3-small", "input": "spend probe"},
+                request_method="POST",
+            ),
+            "litellm_params": {},
+        }
+        result = OpenAIPassthroughLoggingHandler.openai_passthrough_handler(
+            httpx_response=self._create_mock_httpx_response(response_body),
+            response_body=response_body,
+            logging_obj=self._create_mock_logging_obj(),
+            url_route="https://api.openai.com/v1/embeddings",
+            result="",
+            start_time=self.start_time,
+            end_time=self.end_time,
+            cache_hit=False,
+            request_body={"model": "text-embedding-3-small", "input": "spend probe"},
+            **kwargs_in,
+        )
+
+        assert result["result"] is None
+        assert result["kwargs"]["passthrough_logging_payload"] == kwargs_in["passthrough_logging_payload"]
+
     @patch(
         "litellm.proxy.pass_through_endpoints.llm_provider_handlers.openai_passthrough_logging_handler.OpenAIPassthroughLoggingHandler.openai_passthrough_handler"
     )