From f44b3c9ce0e64cc13b8b9bd30a1762e67f496d1a Mon Sep 17 00:00:00 2001 From: Clay <5866903+jcd1991@users.noreply.github.com> Date: Sat, 21 Nov 2026 19:25:42 -0500 Subject: [PATCH] fix(langchain): record provider-reported costs --- langfuse/langchain/CallbackHandler.py | 54 ++++++++++++++++++++++++--- tests/unit/test_langchain.py | 45 ++++++++++++++++++++++ 2 files changed, 94 insertions(+), 5 deletions(-) diff --git a/langfuse/langchain/CallbackHandler.py b/langfuse/langchain/CallbackHandler.py index dac76f4fc..ac3f9869e 100644 --- a/langfuse/langchain/CallbackHandler.py +++ b/langfuse/langchain/CallbackHandler.py @@ -1328,6 +1328,7 @@ def on_llm_end( ) llm_usage = _parse_usage(response) + cost_details = _parse_cost(response) # e.g. azure returns the model name in the response model = _parse_model(response) @@ -1338,6 +1339,7 @@ def on_llm_end( generation.update( output=extracted_response, usage_details=llm_usage, + cost_details=cost_details, input=kwargs.get("inputs"), model=model, ).end() @@ -1765,11 +1767,50 @@ def _parse_usage_model(usage: Union[pydantic.BaseModel, dict]) -> Any: "cache_creation_input_tokens", total_cache_creation_tokens ) + usage_model.pop("cost", None) usage_model = {k: v for k, v in usage_model.items() if isinstance(v, int)} return usage_model if usage_model else None +def _parse_cost(response: LLMResult) -> Any: + def parse_usage_cost(usage: Any) -> Any: + if hasattr(usage, "__dict__"): + usage = usage.__dict__ + + if not isinstance(usage, dict): + return None + + cost = usage.get("cost") + if isinstance(cost, (int, float)) and not isinstance(cost, bool): + return {"total": cost} + + return None + + if response.llm_output is not None: + for key in ["token_usage", "usage"]: + if key in response.llm_output and response.llm_output[key]: + cost_details = parse_usage_cost(response.llm_output[key]) + if cost_details is not None: + return cost_details + + for generation in response.generations: + for generation_chunk in generation: + response_metadata = getattr( + getattr(generation_chunk, "message", None), "response_metadata", None + ) + if not isinstance(response_metadata, dict): + continue + + for key in ["token_usage", "usage"]: + if key in response_metadata: + cost_details = parse_usage_cost(response_metadata[key]) + if cost_details is not None: + return cost_details + + return None + + def _parse_usage(response: LLMResult) -> Any: # langchain-anthropic uses the usage field llm_usage_keys = ["token_usage", "usage"] @@ -1798,18 +1839,21 @@ def _parse_usage(response: LLMResult) -> Any: chunk_usage = ( ( - response_metadata.get("usage", None) # for Bedrock-Anthropic + response_metadata.get("token_usage", None) + if isinstance(response_metadata, dict) + else None + ) + or ( + response_metadata.get("usage", None) if isinstance(response_metadata, dict) else None ) or ( - response_metadata.get( - "amazon-bedrock-invocationMetrics", None - ) # for Bedrock-Titan + response_metadata.get("amazon-bedrock-invocationMetrics", None) if isinstance(response_metadata, dict) else None ) - or getattr(message_chunk, "usage_metadata", None) # for Ollama + or getattr(message_chunk, "usage_metadata", None) ) if chunk_usage: diff --git a/tests/unit/test_langchain.py b/tests/unit/test_langchain.py index c55eaa3d2..895226e82 100644 --- a/tests/unit/test_langchain.py +++ b/tests/unit/test_langchain.py @@ -172,6 +172,51 @@ def test_llm_callback_exports_generation_span(langfuse_memory_client, get_span): ) +@pytest.mark.parametrize("usage_source", ["llm_output", "response_metadata"]) +@pytest.mark.parametrize("reported_cost", [0.000138, 1]) +def test_chat_model_callback_records_provider_reported_cost( + langfuse_memory_client, get_span, json_attr, usage_source, reported_cost +): + usage = { + "prompt_tokens": 8, + "completion_tokens": 93, + "total_tokens": 101, + "cost": reported_cost, + } + response = ChatResult( + generations=[ + ChatGeneration( + message=AIMessage( + content="bonjour", + response_metadata=( + {"token_usage": usage} + if usage_source == "response_metadata" + else {} + ), + ) + ) + ], + llm_output={"token_usage": usage} if usage_source == "llm_output" else None, + ) + + with patch.object(ChatOpenAI, "_generate", return_value=response): + ChatOpenAI(api_key="test", model="openai/gpt-5-mini").invoke( + "hello", config={"callbacks": [CallbackHandler()]} + ) + + langfuse_memory_client.flush() + span = get_span("ChatOpenAI") + + assert json_attr(span, LangfuseOtelSpanAttributes.OBSERVATION_COST_DETAILS) == { + "total": reported_cost + } + assert json_attr(span, LangfuseOtelSpanAttributes.OBSERVATION_USAGE_DETAILS) == { + "input": 8, + "output": 93, + "total": 101, + } + + def test_lcel_chain_exports_intermediate_chain_spans( langfuse_memory_client, get_span, find_spans ):