Introduction

If automatic instrumentation of LLM requests fails, you can manually create traces and submit completions. This is useful for debugging purposes or to customize your traces and completions.

Note: This functionality is currently only supported in the Python SDK

Usage

There are four steps to manually create traces and submit completions:

  1. Start a manual trace using baserun.start_manual_trace
  2. Call OpenAI to get your completion as normal
  3. Submit a completion using baserun.submit_openai_completion using the arguments passed to OpenAI and then completion returned by OpenAI
  4. Finish the trace using baserun.finish_manual_trace

Example

client = OpenAI()

trace = baserun.start_manual_trace("Manual Submission Test")

openai_kwargs = {
"model": "gpt-4-1106-preview",
"messages": [{"role": "user", "content": "What is the capital of the US?"}]
}
completion = client.chat.completions.create(**openai_kwargs)
baserun.submit_openai_completion(completion, trace, **openai_kwargs)

# You can manually set the result of the trace, or any other metadata on the trace
trace.result = completion.choices[0].message.content
baserun.finish_manual_trace(trace)