feature(wren-ai-service): integrate Langfuse SDK to represent the evaluation result

[paopa]

This PR aims to integrate the decorator-based Langfuse SDK to collect data for the evaluation process. By default, the environment setup is disabled. To enable Langfuse, follow these steps:

# .env.dev file
# Langfuse configuration
LANGFUSE_ENABLE=True
LANGFUSE_SECRET_KEY=
LANGFUSE_PUBLIC_KEY=
LANGFUSE_HOST=

Fill in the environment variables. You can generate the secret key, public key, and host by following the instructions at Langfuse Documentation.

Currently, we don't set the session ID. When implementing the main process for the evaluation, refer to the following code to set up the session ID.

@observe()
async def main(user_id: str):
    langfuse_context.update_current_trace(
        user_id=user_id, 
        session_id=f"{user_id}_{uuid.uuid4()}"
    )
    return await story()


async def run():
    await asyncio.gather(main("foo"), main("bar"))
    langfuse_context.flush()


asyncio.run(run())

Screenshots

[cyyeh]

should we call langfuse_context.flush() during the server shutdown event in the lifespan function in wren-ai-service/src/__main__.py?

Also in each pipeline, you call langfuse_context.flush() in the end, I wonder what would happen if langfuse is disabled then?

[cyyeh]

I suppose we also need to add langfuse related env to .env.prod.example, wren-ai-service/docker/docker-compose.yml, docker/.env.example and docker/docker-compose-dev.yaml, docker/docker-compose.yaml?

[cyyeh]

Maybe we can also setup thie environment variable in the future? LANGFUSE_DEBUG

[cyyeh]

This can be added in the initialization part: Authentication check

[cyyeh]

We should manually add input/output data if needed: https://langfuse.com/docs/sdk/python/decorators#large-inputoutput-data

[cyyeh]

Overall LGTM, I would like to discuss with you the capture_input part

[paopa]

After a discussion with @cyyeh, we decided not to capture the span input for all steps. Other suggestions, such as the Langfuse debug mode and authentication check, are interesting features. However, we will not include them in this PR as it is focused on the evaluation framework.

[cyyeh]

LGTM