perf: avoid quadratic start check when streaming from OpenAIResponsesChatGenerator - #13716

Related Issues

• No existing issue.

Proposed Changes:

_convert_response_chunk_to_streaming_chunk sets StreamingChunk.start on each response.output_text.delta by checking whether an earlier chunk with the same output_index already has content. On main it walks previous_chunks from the front for every delta. When a reasoning summary streams before the answer, each text delta first walks past every reasoning chunk, so a stream with R reasoning deltas and T text deltas does about R x T checks.

The check now walks reversed(previous_chunks) with the same predicate, all(c.content == "" ...), and stops at the previous text delta, which is usually the last chunk. The separate all(c.index != chunk.output_index ...) clause goes away because all() over no matching chunks is already True. The refusal check gets the same reversed(); any() gives the same result in either order. No new state. run_async and AzureOpenAIResponsesChatGenerator use the same function.

How did you test it?

• pytest -m "not integration" on test/components/generators/, test/components/agents/, test_query_expander.py, test_llm_ranker.py and test/dataclasses/test_streaming_chunk.py: 870 passed. ruff check, ruff format --check and mypy on the module are clean.
• Equivalence: the same ResponseStreamEvent sequences were run through the function on main and on this branch, comparing every StreamingChunk field except the received_at timestamp. 20,000 random sequences (text, refusal, reasoning and function call deltas, items added and done, interleaved output_index values, empty deltas) and full realistic streams gave identical chunks.
• Benchmark: realistic streams (reasoning summary item, then the message), main and this branch in one process, alternating order. "Conversion" calls the function over the whole stream. run() and run_async() use a streaming callback and an httpx.MockTransport that serves the SSE body, so SDK parsing is included. Median of 21 runs, min-max in brackets, Python 3.12, openai 3.49.0, shared Linux machine.

Notes for the reviewer

Answers without reasoning barely change, because main already stops early there. The gain grows with the length of the reasoning that precedes the answer. No test on main pins the start flag of refusal deltas; none was added, since that line only gains reversed(). The pre-commit hooks were not run as such; ruff, mypy, codespell and scripts/release_note_backticks.py were run directly on the changed files.

This PR was fully generated with Claude Code, which also ran the tests and benchmarks above.

Checklist

• I have read the contributors guidelines and the code of conduct.
• I have updated the related issue with new insights and changes.
• I have added unit tests and updated the docstrings.
• I've used one of the conventional commit types for my PR title: fix:, feat:, build:, chore:, ci:, docs:, style:, refactor:, perf:, test: and added ! in case the PR includes breaking changes.
• I have documented my code.
• I have added a release note file, following the contributors guidelines.
• I have run pre-commit hooks and fixed any issue.