Trace LLM workflows at your app's semantic level, not at the OpenAI API boundary