Summary
LiteLLM provides some LangFuse support but it's subpar comparing to Typescript counterparts (like @langfuse/vercel-ai-sdk). Being able to pick up from the environment and trace LiteLLM calls that includes execution ids, input and result payloads, contexts, etc would be a nice additiona to observability.
Proposed solution
Alternatives considered
Manually adding headers (ie: x-litellm-session-id) to every acompletion call
Additional context
If it makes sense to set the LangFuse observability on the control plane instead of agents would be even better, but then some additional payloads would need to be sent to the control plane to keep track of internal LiteLLM calls
Summary
LiteLLM provides some LangFuse support but it's subpar comparing to Typescript counterparts (like
@langfuse/vercel-ai-sdk). Being able to pick up from the environment and trace LiteLLM calls that includes execution ids, input and result payloads, contexts, etc would be a nice additiona to observability.Proposed solution
Alternatives considered
Manually adding headers (ie:
x-litellm-session-id) to everyacompletioncallAdditional context
If it makes sense to set the LangFuse observability on the control plane instead of agents would be even better, but then some additional payloads would need to be sent to the control plane to keep track of internal LiteLLM calls