fix: enhance llama-index callback support for exception events

[axiomofjoy]

LlamaIndex recently changed the way their callback system handles exception information. Previously, exceptions always resulted in independent exception callback events. Now, whenever possible, exception information is passed via the on_event_end callback hook for the callback event that caused the exception. When that is not possible (e.g., when the exception arises outside of any event), the callback system produces an independent exception event containing the exception information. This latter scenario should be rare and indicates that there is a bug in LlamaIndex itself.

This PR adjusts our callback handler to accommodate these changes. In particular, it searches each callback event payload for an EventPayload.EXCEPTION key, and if present, grabs the exception information from that key. In the infrequent but possible event that the LlamaIndex callback system registers an independent exception callback event, we create a span of span kind "chain" and attach the exception information to this span.

Note that LlamaIndex is currently incorrectly passing the exception information for failed LLM calls (e.g., due to rate limit errors) to the parent synthesis span in the case of query engines. I've adjusted the unit tests accordingly for now. I've noted to the LlamaIndex team here.

I will handle percolating exceptions up the span tree in a subsequent PR.

Relevant issues:

• https://github.com/Arize-ai/phoenix/pull/1814/files
• [Bug]: exception events are inconsistently placed in the callback trace map run-llama/llama_index#9070

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[RogerHYang]

should we bump the LLAMA_INDEX_MINIMUM_VERSION_TRIPLET?

[axiomofjoy]

should we bump the LLAMA_INDEX_MINIMUM_VERSION_TRIPLET?

Good catch, addressed in e8c8a71.