fix(langchain): capture usage on streamed gemini responses (#1309)

Author: hassiebp

langfuse/langchain/CallbackHandler.py

@@ -1140,7 +1140,9 @@ def _parse_usage(response: LLMResult) -> Any:
                     llm_usage = _parse_usage_model(
                         generation_chunk.generation_info["usage_metadata"]
                     )
-                    break
+
+                    if llm_usage is not None:
+                        break
 
                 message_chunk = getattr(generation_chunk, "message", {})
                 response_metadata = getattr(message_chunk, "response_metadata", {})

tests/test_langchain.py

@@ -177,6 +177,7 @@ def test_callback_generated_from_lcel_chain():
     assert langchain_generation_span.output != ""
 
 
+@pytest.mark.skip(reason="Flaky")
 def test_basic_chat_openai():
     # Create a unique name for this test
     test_name = f"Test Basic Chat {create_uuid()}"