[nadeem4]

What

structured() and streaming_structured() now treat everything before the parser as the model, so the parse-and-retry wrapper applies to langchain-openai 1.7's three-step chain (model, refusal check, parser).

Why

CI started failing (20 unit tests, test (3.12), test (3.13), integration) after dependencies moved to langchain-core 1.7.5 / langchain-openai 1.7.4. It is not pydantic 2.14: ValidationError is still a ValueError there. langchain-openai 1.7 inserts a _check_refusal step between the model and the parser, so _split (which required exactly two steps) returned None, the retry wrapper was skipped, and a bad answer's ValidationError escaped instead of being retried. The extra steps run directly under the caller's run (not as a nested sequence) so the trace still attaches the parsed answer to the model call.

How to test

pytest packages/nl2sql/tests/unit/test_parse_and_retry.py test_chat_tools.py test_conversations.py test_engine_events.py test_cost_and_caps.py test_limits_and_join_keys.py test_llm_providers_v2.py. Passes on pydantic 2.14.2 + langchain-openai 1.7.4 and on pydantic 2.13.5 + langchain-openai 1.6.2.

Checklist

• Failing tests first (the 20 CI failures reproduced locally)
• Boundary tests pass
• No doc change needed (behaviour unchanged)