src/structured.rs

This file holds the types and helpers you use to make your model answer in JSON you can parse. Read it before you mix structured output with tool calls, because the two compete.

Key items

Your StructuredOutput trait gives you name(), schema(), and from_value(), where serde supplies the default.

ResponseFormat { name, schema, strict } plus from_type::<T>() builds a format, and strict defaults to true.

RequestOptions { response_format, tool_constraint, model } carries your per-request switches through its builders, and with_model ignores empty names.

ToolConstraint offers None, Strict, or Grammar(provider), which you reach under the grammar feature.

ThinkingEffort plus RequestOptions::effort(...) (0.4.0) gives you Off, Low, Medium, High, and Max for reasoning models, and RequestOptions stays #[non_exhaustive].

request_structured::<T>(client, messages, system) is your one-shot helper, and it runs with no parse-failure retry.

StructuredError { Deserialize, Api } is what your caller matches on. Internally you get parse_json_lenient and extract_json_substring for the lenient extractor, plus tighten_json_schema for the strict-mode transform.

Behavior notes

A response_format suppresses tools for that request, because the two share one slot.

strict: true formats are rejected up front on Anthropic and Gemini, which have no strict switch, so you never eat a silent downgrade.

Clients that cannot honor options reject loudly, which is the default *_with_options behavior.

Tightening recurses through nesting and $defs, closes every object, and makes every property required. It is idempotent, and it is exactly what OpenAI strict mode requires.

Deep dive: Structured output.