fix!: align chat completions API with OpenAI spec

[cdoern]

What does this PR do?

• Fix GET /chat/completions query parameters to match OpenAI:
  • Remove nullable types from after, limit, model, order
  • Add enum constraint to order parameter (asc, desc)
  • Set correct default values
• Make finish_reason nullable in OpenAIChoice and OpenAIChunkChoice to match OpenAI streaming response spec
• Make input_messages optional in OpenAICompletionWithInputMessages since it's not part of standard OpenAI response Fixed errors:

  error    [request-parameter-became-enum] at docs/static/openai-spec-2.3.0.yml
      in API GET /chat/completions
          the 'query' request parameter 'order' was restricted to a list of enum values

  error    [request-parameter-default-value-changed] at docs/static/openai-spec-2.3.0.yml
      in API GET /chat/completions
          for the 'query' request parameter 'order', default value was changed from 'desc' to 'asc'

  error    [request-parameter-list-of-types-narrowed] at docs/static/openai-spec-2.3.0.yml
      in API GET /chat/completions
          'query' request parameter 'after' list-of-types was narrowed by removing types 'null'

  error    [request-parameter-list-of-types-narrowed] at docs/static/openai-spec-2.3.0.yml
      in API GET /chat/completions
          'query' request parameter 'limit' list-of-types was narrowed by removing types 'null'

  error    [request-parameter-list-of-types-narrowed] at docs/static/openai-spec-2.3.0.yml
      in API GET /chat/completions
          'query' request parameter 'model' list-of-types was narrowed by removing types 'null'

  error    [request-parameter-list-of-types-narrowed] at docs/static/openai-spec-2.3.0.yml
      in API GET /chat/completions
          'query' request parameter 'order' list-of-types was narrowed by removing types 'null'

  error    [request-parameter-type-changed] at docs/static/openai-spec-2.3.0.yml
      in API GET /chat/completions
          for the 'query' request parameter 'limit', the type/format was changed from ''/'' to 'integer'/''

  error    [response-required-property-removed] at docs/static/openai-spec-2.3.0.yml
      in API GET /chat/completions
          removed the required property 'data/items/input_messages' from the response with the '200' status

  error    [response-required-property-removed] at docs/static/openai-spec-2.3.0.yml
      in API GET /chat/completions/{completion_id}
          removed the required property 'input_messages' from the response with the '200' status

The changes here align the GET /chat/completions endpoint with OpenAI's API spec.

When you write after: str | None = None, the schema generator produces:

anyOf:                                                                                                                                                                        
  - type: string                                                                                                                                                              
  - type: 'null'                                                                                                                                                              

But OpenAI's spec has these parameters as optional but not nullable — you can omit them, but you can't explicitly pass null.

Using after: str = None produces the correct schema:
type: string

The # type: ignore[assignment] silences mypy since None isn't technically a valid str, but the schema generator only sees str and outputs the correct OpenAPI spec.

I looked at adopting the _remove_null_from_anyof approach from #4644, but that works for Pydantic model fields (like the embeddings request body), not for query parameters on @webmethod decorated methods. The custom schema generator doesn't process Query(json_schema_extra=...) metadata. Once Inference is converted to a FastAPI router, we could revisit this.

resolves #4622

Test Plan

conformance diff between LLS spec and openAI spec should not have these errors anymore.

[eoinfennessy]

I've proposed a way of allowing for optional but non-nullable types in #4644.

Specifically, I created a helper _remove_null_from_anyof to ensure null does not get added to the OpenAPI schema, and I added runtime validation to ensure null is not explicitly provided.

This allows the type annotations to remain accurate (e.g. str | None with no MyPy errors) while ensuring the spec & runtime behavior conforms to the OpenAI spec and the runtime.

Would be great to get your thoughts on this.

[cdoern]

thanks! I took a look, and compared. I am going to adopt your _remove_null_from_anyof here. Both of our approaches work, but yours is more sustainable and reusable

[cdoern]

actually -- I am seeing issues here. the inference API is still using web method not fastAPI routers, this causes the openapi spec still to have anyOf if I use your approach. might have to stick with this until the inference API is converted.

[mattf]

we need to avoid type ignoring

[mattf]

this change requires updates to providers, which are using order.value

[leseb]

another example of why we need to use request classes so we don't have to update all providers impls... soon with #4445

[cdoern]

yeah. I can make this more proper. Can hold off until #4445 merges.

[leseb]

that doesn't seem to be correct, the spec says:

finish_reason:
  type: string
  description: >
    The reason the model stopped generating tokens. This will be `stop` if the model hit a
    natural stop point or a provided stop sequence,

    `length` if the maximum number of tokens specified in the request was reached,

    `content_filter` if content was omitted due to a flag from our content filters,

    `tool_calls` if the model called a tool, or `function_call` (deprecated) if the model called
    a function.
  enum:
    - stop
    - length
    - tool_calls
    - content_filter
    - function_call

Shouldn't this be?

OpenAIFinishReason = Literal["stop", "length", "tool_calls", "content_filter", "function_call"]
@json_schema_type
class OpenAIChunkChoice(BaseModel):
...
...
    finish_reason: OpenAIFinishReason | None = None
...
...


@json_schema_type
class OpenAIChoice(BaseModel):
...
...
    finish_reason: OpenAIFinishReason
...
...

[cdoern]

putting in draft waiting for a few PRs like the fastAPI migration of inference + the new conformance test to merge. Will revisit then!