Schemas as OpenAI structured output (Effect's
effect/ai/OpenAiStructuredOutput).
to_codec_openai(schema) gives the JSON Schema to send to OpenAI's
strict mode and the codec decoding the model's output back into the
schema's values. Shapes strict mode cannot express are sent in the
provider-safe form anthropic_structured_output describes (tuples as
objects, records as [key, value] pairs, optional keys as required
nullable ones, one_of as anyOf). OpenAI takes no allOf: patterns
of its members are merged into one, bounds into the tightest, and the
other keywords kept only when that keeps their meaning. Keywords OpenAI
does not support, or that do not apply to the schema's type, are left out
(an unsupported format becomes part of the description) and still
checked by the codec. The root must be an object without anyOf.
Compatibility targets standard models; fine-tuned ones take less.
Classes#
Functions#
to_codec_openai#
to_codec_openai[T, E, RD, RE](schema) -> StructuredOutputCodec[T, RD, RE]schema's OpenAI JSON Schema and the codec of its output (see the
module); raises ValueError when the root is not an object.
| Parameter | Type | Description |
|---|---|---|
schema | Codec[T, E, RD, RE] |
Returns StructuredOutputCodec[T, RD, RE]
def to_codec_openai[T, E, RD, RE](
schema: Codec[T, E, RD, RE], /
) -> StructuredOutputCodec[T, RD, RE]:
"""``schema``'s OpenAI JSON Schema and the codec of its output (see the
module); raises ``ValueError`` when the root is not an object."""
codec = to_codec(schema)
document = J.document(codec)
root = _rewrite(J.resolve_root(document))
if root.get("type") != "object" or "anyOf" in root:
raise ValueError(
'OpenAiStructuredOutput: Root JSON Schema must have type "object" '
'and must not use "anyOf"'
)
return StructuredOutputCodec(codec, J.with_definitions(root, document.definitions, _rewrite))