[HN Gopher] Segment for LLM Traces? Seeking Feedback on an Open ...
___________________________________________________________________
Segment for LLM Traces? Seeking Feedback on an Open Source LLM Log
Router
Hey everyone, I'm considering starting a new open source project
and wanted to see if anyone else thinks the idea could be useful.
The concept is simple: an open source LLM log router that works
like Segment--but specifically for LLM logs. It would let you
easily route logs to different analytics, eval, monitoring, and
data warehouse platforms (think LangFuse, Gentrace, Lattitude,
etc.) so you can leverage the strengths of each without needing
separate integrations. I've run into a few recurring challenges
when integrating multiple eval and monitoring tools. Conflicting
Integrations: Many tools use their own forks of popular packages
(like the OpenAI SDK or LangChain), which often conflict--making it
nearly impossible to use them together. Inconsistent Prompt
Templating: Sometimes different tools require different prompt
formats, complicating the process of switching between them or
using multiple tools simultaneously. Data Migration Challenges:
Moving logs between systems is a hassle. Testing a new tool often
means generating new data or deploying changes in production,
making it hard to evaluate if switching is worthwhile. While some
solutions exist (such as LangChain and LiteLLM integrations with
various eval platforms), they don't fully address these issues--
especially data migration and integrating with multiple tools at
once. To solve this, I'm toying with the idea of a new open source
project--a lightweight, self-hosted server that acts as a "Segment
for LLM traces and logs". Here are some of the features I envision.
Self-hosted logging server: Send LLM traces to a single endpoint
without impacting your app's performance. Centralized aggregation
and routing: Gather traces and logs on a single server, then
forward them to any destination you choose (evals, analytics,
monitoring, alerting, data warehouses, etc). Lightweight Framework
integrations: Support for popular frameworks and SDKs (LangChain,
LiteLLM, OpenAI SDK, LlamaIndex, etc.) with integrations that never
block your event loop--so even if the logging server or a
destination platform goes down your app continues to function as
expected. Easy configuration: A simple interface for managing data
sources and destinations without making code changes or redeploying
your application. Data portability: Store logs in a database with
the option to re-export to new tools in the future--ensuring you
never face vendor lock-in. Custom integrations: Webhooks to easily
set up your own custom destinations. I'd love to hear if you've
experienced any similar issues integrating with multiple LLM
monitoring, eval, and analytics platforms. If so, how did you
address them? Do you see value in an open source data router like
this? Please share your thoughts in the comments, and if you're
interested in contributing or using a project like this in the
future, it'd be super helpful if you could fill out this survey.
https://yk1m5yevl9j.typeform.com/to/cQdxF6bN Thanks in advance for
your feedback!
Author : patethegreat
Score : 4 points
Date : 2025-02-21 22:44 UTC (16 minutes ago)
___________________________________________________________________
(page generated 2025-02-21 23:01 UTC)