Respan Mistral AI instrumentation

Trace the official mistralai Python SDK with Respan.

This package wraps openinference-instrumentation-mistralai and registers Respan's OpenInference translator so Mistral AI spans are emitted with the canonical traceloop.*, gen_ai.*, llm.*, and respan.* fields expected by the Respan OTLP pipeline. It covers synchronous and asynchronous chat, streaming content and usage, request tool definitions, current-turn tool calls, precise provider error status, and interrupted-stream finalization. Duplicate native Mistral SDK spans are suppressed only for the SDK's own instrumentation scope. Tested dependency support starts at mistralai==2.9.3 and openinference-instrumentation-mistralai==2.0.6.

Install

pip install respan-ai respan-instrumentation-mistralai mistralai

Usage

import os

from mistralai.client import Mistral
from respan import Respan
from respan_instrumentation_mistralai import MistralAIInstrumentor

respan = Respan(
    api_key=os.environ["RESPAN_API_KEY"],
    instrumentations=[MistralAIInstrumentor()],
)

with Mistral(api_key=os.environ["MISTRAL_API_KEY"]) as client:
    response = client.chat.complete(
        model="mistral-large-latest",
        messages=[
            {
                "role": "user",
                "content": "Reply with one concise sentence about tracing.",
            }
        ],
    )
    print(response.choices[0].message.content)

respan.flush()
respan.shutdown()

Any keyword arguments passed to MistralAIInstrumentor(...) are forwarded to the underlying OpenInference instrumentor. Multiple Respan Mistral instrumentor instances share one OpenInference delegate and one normalization processor; the final owner deactivates them. Shared instances must use identical keyword arguments. An activation with different arguments is rejected and remains inactive rather than silently inheriting another instance's configuration.