Wrap OpenAI LLM calls with the Langfuse Python drop-in integration to auto-capture tokens, cost, model params, and latency

Documented steps

1. Install both packages: `pip install langfuse openai`.
2. Set env vars `LANGFUSE_PUBLIC_KEY`, `LANGFUSE_SECRET_KEY`, and `LANGFUSE_BASE_URL` exactly as in the base SDK setup.
3. Swap your import for the drop-in module. No other code changes are required: `from langfuse.openai import openai`, or import specific classes with `from langfuse.openai import OpenAI, AsyncOpenAI, AzureOpenAI, AsyncAzureOpenAI`.
4. Call the client exactly as you would the normal OpenAI SDK: `result = openai.chat.completions.create(model="gpt-3.5-turbo", messages=[{"role": "user", "content": "Hello"}], name="my-chat")`. The integration transparently captures prompts/completions (including streaming, async, and function/tool calls), response latency, API errors with level and statusMessage, token usage, USD cost, and model parameters.
5. Nest LLM calls inside a larger trace by combining with the decorator: put `@observe()` on an outer function and call `openai.chat.completions.create(...)` from inside it. The generation becomes a child observation of that trace automatically.
6. Attach session, user, and tag attribution per call using `langfuse_`-prefixed keys in the `metadata` dict: `metadata={"langfuse_session_id": "session_123", "langfuse_user_id": "user_456", "langfuse_tags": ["calculator"], "someMetadataKey": "someValue"}`.
7. Enable verbose debug logging while validating a new integration: `Langfuse(debug=True)`.
8. Control volume and cost with `sample_rate` on the `Langfuse(...)` constructor (or the `LANGFUSE_SAMPLE_RATE` env var) if you only want a fraction of calls traced.
9. Call `langfuse.flush()` before a short-lived process exits; this integration uses the same asynchronous background exporter as the base SDK.
10. Verified against current official docs on 2026-07-29: https://langfuse.com/integrations/model-providers/openai-py