WIP: fix(huggingface): add Pydantic structured output support (Issue #32197)

[rahulmansharamani14]

This draft PR begins addressing #32197 (Support structured output with Pydantic models in langchain-huggingface).

Currently, ChatHuggingFace.with_structured_output converts Pydantic schemas into JSON Schema and parses into dicts. PydanticOutputParser is not utilized, leading to dict outputs and extra token usage on retries.

What’s included

• Detects Pydantic model classes (is_basemodel_subclass).
• Switches to PydanticOutputParser when schema is Pydantic and method="json_schema".
• Preserves existing behavior for JSON Schema dicts, TypedDicts, and json_mode.
• function_calling with Pydantic continues to raise NotImplementedError.

Next steps (planned)

• Add unit tests under libs/partners/huggingface/tests/unit_tests/ for:
  • Round-trip returning Pydantic instances.
  • include_raw=True path.
  • No extra model calls on parse retry.
  • Non-Pydantic schemas unchanged.
• Un-xfail Pydantic-related integration tests once validated.

Notes

Opening early for review to validate direction before completing test coverage and unskipping integration tests.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Preview	Comments	Updated (UTC)
langchain	Ignored	Preview		Sep 28, 2025 9:57pm

[codspeed-hq]

CodSpeed Instrumentation Performance Report

Merging #33141 will create unknown performance changes

_{Comparing rahulmansharamani14:fix/hf-pydantic-structured-output (585b25d) with master (9863023)¹}

Summary

⚠️ No benchmarks were detected in both the base of the PR and the PR.
Please ensure that your benchmarks are correctly instrumented with CodSpeed.

Check out the benchmarks creation guide
⏩ 21 skipped²

Footnotes

1. No successful run was found on master (54ea620) during the generation of this report, so 9863023 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

2. 21 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩