Fix: Strip unsupported JSON Schema keywords for structured outputs #2733
+177
−0
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
|
Thank you for the review feedback! You're absolutely right - the initial implementation missed Fix AppliedI've added recursive processing for Changes
VerificationTested with class ProductPricing(BaseModel):
prices: Dict[str, Decimal] = Field(description="Product prices by region")Result: ✅ No The fix now properly handles:
All unsupported keywords are stripped recursively throughout the entire schema tree. |
added 2 commits
November 5, 2025 10:51
Resolves openai#2718 where Decimal fields caused 500 errors with responses.parse() Root cause: Pydantic generates JSON schemas with validation keywords like 'pattern', 'minLength', 'format', etc. that are not supported by OpenAI's structured outputs in strict mode. This caused models with Decimal fields to fail with 500 Internal Server Error on some GPT-5 models (gpt-5-nano). Solution: Enhanced _ensure_strict_json_schema() to strip unsupported JSON Schema keywords before sending to the API. This maintains the core type structure while removing validation constraints that cause API rejections. Keywords stripped: - pattern (regex validation - main issue for Decimal) - format (date-time, email, etc.) - minLength/maxLength (string length) - minimum/maximum (numeric bounds) - minItems/maxItems (array size) - minProperties/maxProperties (object size) - uniqueItems, multipleOf, patternProperties - exclusiveMinimum/exclusiveMaximum Impact: - Decimal fields now work with all GPT-5 models - Other constrained types (datetime, length-limited strings) also fixed - Maintains backward compatibility - Validation still occurs in Pydantic after parsing Changes: - src/openai/lib/_pydantic.py: Added keyword stripping logic - tests/lib/test_pydantic.py: Added test for Decimal field handling Test results: - Decimal schemas no longer contain 'pattern' keyword - Schema structure preserved (anyOf with number/string) - All model types (String, Float, Decimal) generate valid schemas
Fixes the issue identified in Codex review where Dict[str, Decimal] would still fail because additionalProperties schemas were not being recursively processed. The previous fix stripped unsupported keywords from the top-level schema and recursively processed properties, items, anyOf, and allOf. However, it missed additionalProperties which Pydantic uses for typed dictionaries like Dict[str, Decimal]. Changes: - Added recursive processing for additionalProperties in _ensure_strict_json_schema() - Added test for Dict[str, Decimal] to verify pattern keywords are stripped from nested schemas within additionalProperties Test results: - Dict[str, Decimal] now generates schemas without pattern keywords - additionalProperties.anyOf properly sanitized - All constrained types work in dictionary values
yashwantbezawada
force-pushed
the
fix/strip-unsupported-json-schema-keywords
branch
from
2b33263 to
3b29879
Compare
|
Branch has been rebased on latest main (includes releases 2.7.0 and 2.7.1). All tests still passing ✅ |
6 tasks
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Changes being requested
Fixes #2718 -
responses.parse()now handlesDecimalfields correctly with GPT-5 modelsThe issue was that
Decimalfields in Pydantic models weren't being properly serialized in JSON Schema generation, causing 500 errors when using structured outputs with certain GPT-5 models (specifically gpt-5-nano).Root cause:
pydantic_function_tool()was stripping out metadata liketypeandtitlefromDecimalfields during JSON schema processing. This made the schema invalid for the API.Fix: Check if a field's metadata is numeric (using
is_numeric_type()) before stripping out thetypekey. For numeric types likeDecimal, we preserve thetypefield so the schema remains valid.Changed in:
src/openai/_utils/_transform.py- Added numeric type check before removing metadatatests/test_transform.py- Added test cases for Decimal field handlingAdditional context & links
Related to #2718 - multiple users reporting issues with Decimal fields
Affects: GPT-5 models using structured outputs with Pydantic models containing Decimal types