fix(llama-index): Suppress unhandled event warnings for SparseEmbeddingEndEvent and Workflow*OutputEvent

[enrico-stauss]

Fix LlamaIndex and LlamaIndex Workflows Unhandled Event Logs

This PR registers singledispatch handlers for four previously unhandled event types that were producing spurious WARNING logs:

• WARNING - Unhandled event of type SparseEmbeddingEndEvent
• WARNING - Unhandled event of type WorkflowStepOutputEvent
• WARNING - Unhandled event of type WorkflowRunOutputEvent
• WARNING - Unhandled event of type SpanCancelledEvent (pre-emptive)

The fix will take effect once the companion fixes in the llama_index and workflows-py repositories are also merged.

llama-index-core -> SparseEmbeddingEndEvent

Find the PR here.
Sparse embedding vectors are Dict[int, float] (token-id → weight), incompatible with the EMBEDDING_VECTOR attribute (List[float]).
The handler records the text chunks and skips the vectors, keeping the span populated without violating the semantic convention.
Note: SparseEmbeddingStartEvent no longer needs a handler here because a companion fix in llama-index-core makes it inherit from EmbeddingStartEvent, so singledispatch routes it via MRO.

llama-index-workflows (optional dependency) -> WorkflowStepOutputEvent/WorkflowRunOutputEvent/SpanCancelledEvent

Find the PR here.
These events are emitted by the workflows package (llama-index-workflows). Handlers are registered inside a try/except ImportError block so the instrumentor continues to work when that package is not installed.
WorkflowStepOutputEvent and WorkflowRunOutputEvent carry an output string that is recorded as OUTPUT_VALUE on the span. SpanCancelledEvent is silenced with a no-op.

[enrico-stauss]

I did the follow-up after realizing that self[OUTPUT_VALUE] = event.output serializes pydantic models to just the first characters produced by the __str__ method which is not very informative.

I don't really know how LlamaIndex coordinates this type of change with you guys but I would assume that this requires acceptance from maintainers of the LlamaIndex and LlamaIndex Workflows repos?

Anyway I tested this in my local setup by referencing all 3 PRs. The warnings are gone and the output looks good to me (the only change is in the SparseEmbeddingOutputEvent attributes). I hope this gets merged since the sheer amount of logged warnings was honestly quite annoying.

Cheers
Enrico

[adrianlyjak]

Chiming in from the llamaindex side. Having this change spread across 3 repos seems pretty hard to get right and make properly compatible.

Perhaps this default should just be changed to a debug, or have something that more reasonably dumps the event contents as (maybe truncated) json

    @singledispatchmethod
    def _process_event(self, event: BaseEvent) -> None:
        logger.warning(f"Unhandled event of type {event.__class__.__qualname__}")