Summary

Fixes #1821. Resolves repeated PydanticSerializationUnexpectedValue warnings logged for every streamed chat chunk. Langflow's /response SSE always wraps text deltas as {"content": "..."} for every provider, but the openai SDK's response event models declare delta: str. When chunk.model_dump() runs, Pydantic's serializer checks delta against that str type, finds a dict, and logs a warning for every chunk which pollutes the openrag-backend logs

Changes

• exclude delta from model_dump()'s serialization (which is what triggers the type-check warning), then reattach the raw, unvalidated delta value back onto the resulting dict. Output shape sent to the frontend/SDK is unchanged and only the noisy serializer warning is avoided.
• Added a comment above my agent.py changes explaining why and when this workaround can be removed.
• Added a simple regression unit test to verify that the pydantic warning isnt being raised and the fields to be expected