langchain: wrap ChatMistralAI

[codefromthecrypt]

If I run the following with EDOT python+Langtrace, I get only an HTTP span, but nothing from Langtrace.

from langchain_mistralai.chat_models import ChatMistralAI

# ChatMistralAI is implemented directly in Langchain using httpx, so install
# the Langchain instrumentation.
from langtrace_python_sdk.instrumentation import LangchainInstrumentation

LangchainInstrumentation().instrument()


def main():
    llm = ChatMistralAI(model_name="open-mistral-nemo", streaming=True)

    response_stream = llm.stream("Write a short poem on OpenTelemetry.")
    for chunk in response_stream:
        print(chunk.content, end="")


if __name__ == "__main__":
    main()

If we had instrumentation here, we would get gen_ai spans for Mistral when used in Langchain.

Notes

• I discussed the idea of re-using the official SDK here, which might work with langtrace's vertex instrumentation if implemented. Use mistralai package instead of raw httpx to implement ChatMistralAI langchain-ai/langchain#29432
• since this is mostly openai compatible code, it could be possible to make a httpx based instrumentation that handles known chat completion expressions. I'm not sure if it would work.

[alizenhom]

Thanks for pointing out the issue,

Will take a look and keep you posted @codefromthecrypt

[codefromthecrypt]

To elaborate one idea I had is if there's a way to handle httpx layer "openai compatible stuff", basically some code could be extracted to a common library and used in different places. I'm not sure how practical this would be.