openai[patch]: support optional fields in dict structured output with method="json_schema"

[ccurme]

The json_schema method will currently error on the following:

Optional fields in TypedDict:

class Entity(TypedDict):
    """Extracted entity."""

    animal: Annotated[str, ..., "The animal"]
    color: Annotated[Optional[str], None, "The color"]

structured_llm = llm.with_structured_output(Entity, method="json_schema")
structured_llm.invoke("black cat")

Optional fields in JSON schema generated from a Pydantic object:

class Entity(BaseModel):
    """Extracted entity."""
    animal: str = Field(description="The animal")
    color: Optional[str] = Field(default=None, description="The color")

structured_llm = llm.with_structured_output(Entity.model_json_schema(), method="json_schema")
structured_llm.invoke("black cat")

Fields with any non-null default (Pydantic, TypedDict, or JSON schema):

class Entity(BaseModel):
    """Extracted entity."""
    animal: str = Field(description="The animal")
    color: str = Field(default="blue", description="The color")

structured_llm = llm.with_structured_output(Entity, method="json_schema")
structured_llm.invoke("black cat")

Here, we update _convert_to_openai_response_format to update optional fields to be required with possible 'null' type. This is a solution recommended in OpenAI's docs.

Note that we make no changes to schemas specified as Pydantic models.

TODO:

• support nested schemas for TypedDict, which also currently raise an error.
• run new standard test across providers

[vercel]

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

1 Skipped Deployment

Name	Status	Preview	Comments	Updated (UTC)
langchain	⬜️ Ignored (Inspect)			Jan 7, 2025 8:01pm