Skip to content
loopctl

`src/structured.rs` — structured (JSON) output

Types and helpers for making the model answer in JSON you can parse.

Key items

  • StructuredOutput trait — name(), schema(), from_value() (default: serde).
  • ResponseFormat { name, schema, strict } + from_type::<T>() (strict default true).
  • RequestOptions { response_format, tool_constraint, model } + builders (with_model ignores empty names).
  • ToolConstraint — None / Strict / Grammar(provider) [grammar feature].
  • request_structured::<T>(client, messages, system) — one-shot helper; no parse-failure retry.
  • StructuredError { Deserialize, Api }.
  • (internal) parse_json_lenient / extract_json_substring — the lenient JSON extractor; tighten_json_schema — the strict-mode schema transform.

Behavior notes

  • A response_format suppresses tools for that request — the two share one slot.
  • strict: true formats are rejected up front on Anthropic/Gemini (no strict switch there) — never a silent downgrade.
  • Clients that can’t honor options reject loudly (default *_with_options behavior).
  • Tightening recurses through nesting and $defs, closes every object, and makes every property required — idempotent, and exactly what OpenAI strict mode requires.

Deep dive: Structured output.