chore: use Pydantic to generate OpenAPI schema

Browse source

Removes the need for the strong_typing and pyopenapi packages and purely
use Pydantic for schema generation.

Our generator now purely relies on Pydantic and FastAPI, it is available
at `scripts/fastapi_generator.py`, you can run it like so:

```
uv run ./scripts/run_openapi_generator.sh
```

The generator will:

* Generate the deprecated, experimental, stable and combined specs
* Validate all the spec it generates against OpenAPI standards

A few changes in the schema required for oasdiff some updates so I've
made the following ignore rules. The new Pydantic-based generator is
likely more correct and follows OpenAPI standards better than the old
pyopenapi generator. Instead of trying to make the new generator match
the old one's quirks, we should focus on what's actually correct
according to OpenAPI standards.

These are non-critical changes:

* response-property-became-nullable: Backward compatible:
  existing non-null values still work, now also accepts null
* response-required-property-removed: oasdiff reports a false
  positive because it doesn't resolve $refs inside anyOf; we could use
  tool like 'redocly' to flatten the schema to a single file.
* response-property-type-changed: properties are still object
  types, but oasdiff doesn't resolve $refs, so it flags the missing
  inline type: object even though the referenced schemas define type:
  object
* request-property-one-of-removed: These are false positives
  caused by schema restructuring (wrapping in anyOf for nullability,
  using -Input variants, or simplifying nested oneOf structures)
  that don't change the actual API contract - the same data types are
  still accepted, just represented differently in the schema.
* request-parameter-enum-value-removed: These are false
  positives caused by oasdiff not resolving $refs - the enum values
  (asc, desc, assistants, batch) are still present in the referenced
  schemas (Order and OpenAIFilePurpose), just represented via schema
  references instead of inline enums.
* request-property-enum-value-removed: this is a false positive caused
    by oasdiff not resolving $refs - the enum values (llm, embedding,
    rerank) are still present in the referenced ModelType schema,
    just represented via schema reference instead of inline enums.
* request-property-type-changed: These are schema quality issues
    where type information is missing (due to Any fallback in dynamic
    model creation), but the API contract remains unchanged -
    properties still exist with correct names and defaults, so the same
    requests will work.
* response-body-type-changed: These are false positives caused
  by schema representation changes (from inferred/empty types to
  explicit $ref schemas, or vice versa) - the actual response types
  an API contract remain unchanged, just how they're represented in the
  OpenAPI spec.
* response-media-type-removed: This is a false positive caused
  by FastAPI's OpenAPI generator not documenting union return types with
  AsyncIterator - the streaming functionality with text/event-stream
  media type still works when stream=True is passed, it's just not
  reflected in the generated OpenAPI spec.
* request-body-type-changed: This is a schema correction - the
  old spec incorrectly represented the request body as an object, but
  the function signature shows chunks: list[Chunk], so the new spec
  correctly shows it as an array, matching the actual API
  implementation.

Signed-off-by: Sébastien Han <seb@redhat.com>

...

This commit is contained in:

Sébastien Han 2025-10-29 14:38:56 +01:00

parent a078f089d9

commit e3cb8ed74a

No known key found for this signature in database

45 changed files with 46606 additions and 18616 deletions

Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL

Download patch file Download diff file Expand all files Collapse all files

14556

docs/static/stainless-llama-stack-spec.yaml vendored

View file

File diff suppressed because it is too large Load diff