The LLM monitoring/evaluations space has seen a number of products off late, both open source and closed source ones. But, a couple of things we have observed are: lack of standard spans and traces that creates vendor lock in, different tools are optimized for solving different pain points - Evaluations, Prompt management, Datasets etc.
We believe that adopting open telemetry(OTEL) standard tracing not only lets teams use our SDK without having to switch their observability client, but will also enable developers to develop tooling for any custom needs such as capturing datasets, prompts, evaluations etc.
A quick note on what we have built so far:
[1] We have a Python and a TypeScript SDK and we have broken down the support for the LLM layer into 3 groups, LLMs, Frameworks and VectorDBs. Our SDKs are open telemetry compatible, can be installed and used independently and we also provide an option to pass custom exporters to export the traces and spans to any observability tool of your choice.
[2] An observability client that is hyper optimized for solving the unique pain points and challenges that come with LLM based apps like Evaluations, prompt iteration, datasets etc. We are SOC2 compliant and the client can also be self hosted if you have strict data privacy and protection requirements.
[3] Both the SDK and the client are fully open source. We are leaning on the community to try it out and provide feedback.
A note about Open Telemetry semantic conventions for LLMs - We would like to converge on standard names for trace attributes that follow the OTEL rules and are looking for feedback from experts here - https://github.com/Scale3-Labs/langtrace/discussions/71
We recognize that this project is early and there is a lot of room for improvement. Would love to hear your thoughts and feedback. Thanks!
Links: