langfuse

Use the langfuse.openai drop-in replacement for the OpenAI client, which automatically traces all calls. Pass extra parameters like name, session_id, user_id, tags, and metadata directly in the completion call, including for streaming and async usage.

Create a Langfuse CallbackHandler with your keys and pass it via config={"callbacks": [handler]} when invoking chains, agents, or retrievers. You can also set it as the default handler so all LangChain calls are traced automatically.

Yes, but traces are batched, so you must call langfuse.flush() before the function exits or data is lost. For critical traces, consider synchronous mode or context managers to guarantee delivery.

Missing data usually comes from not flushing the client before process exit, or from not ending generations with output and usage fields. Always call generation.end() with the response content and token usage, then flush.

Self-hosting requires managing your own infrastructure, and high-volume workloads may need optimization. The real-time dashboard has some latency, and evaluation features require additional setup such as datasets and scoring configuration.