YAshhh29 commented on issue #73658: URL: https://github.com/apache/airflow/issues/73658#issuecomment-5835774878
Tested #71989 against `apache-airflow-providers-openai==2.0.0rc1` from PyPI (Airflow 3.3.2, Python 3.13.3, `openai` 3.19.2), and it works as expected. The merge commit (d9a5f77) is in the RC tag, and the installed files match it byte for byte. Using the real OpenAI SDK with only the HTTP transport mocked, all 10 checks pass, including: - a batch of strings returns one vector per input, in input order, even when the API answers out of order - a batch of token arrays returns one vector per array - a single string or token array still returns one vector - the operator's XCom follows the same shapes - for a batch, `result[0]` equals what 1.8.2 returned Against 1.8.2, the six checks for the fix fail and the four for unchanged behaviour pass. One small, non-blocking suggestion on the 2.0.0 warning: it says the new shape applies "when it is given a list", but a single token array (`list[int]`) is also a list and still returns one vector, so following "take `result[0]`" would give that user a float. Maybe "when it is given a batch", plus "a single string or token array is unchanged"? cc @shahar1 Tested on native Windows (no fork there), with `os.register_at_fork` stubbed during the Airflow imports. No live API calls. Drafted-by: Claude Code (Opus 5.5); reviewed by @YAshhh29 before posting -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
