Arize-ai / phoenix

AI Observability & Evaluation
https://docs.arize.com/phoenix
Other
3.81k stars 283 forks source link

[ENHANCEMENT] Unfurl messages in non-LLM span kinds #4938

Open Jgilhuly opened 3 weeks ago

Jgilhuly commented 3 weeks ago

Describe the Enhancement We currently unfurl input messages for spans with span_kind=LLM. We do not unfurl messages for other span_kinds, like chains and agent spans. Some of our instrumentors, including Langchain, log the same messages object that is sent in an LLM span in other CHAIN spans. Unfurling these messages in the same way as we do for LLMs would make the spans much easier to digest.

Screenshots

Screenshot 2024-10-08 at 10 24 36 AM
Jgilhuly commented 2 weeks ago

I'm not sure how we'd actually do this given inputs have so much variability. Hopefully there's something I'm missing, otherwise we may need to close this one

axiomofjoy commented 2 weeks ago

@Jgilhuly By unfurl, do you mean the list of input and output messages we show for LLM spans? If so, we are able to show that more opinionated UI for LLM spans because we have additional semantic conventions specifically for message content, role, etc., that we don't have for other span kinds.