Source: openaf 06414cd9. See documentation baseline for availability and verification.

OpenAI-compatible schema output

jsonSchemaPrompt(prompt, descriptor) and jsonSchemaPromptWithStats use the OpenAI provider’s structuredSchemaPrompt hook. The descriptor contains name, schema, optional description, and strict (defaults to true).

var llm = $llm({ type: "openai", options: {
  key: "...", url: "https://provider.example/v1", model: "model-name",
  schemaTransport: "chat"
}});
var result = llm.jsonSchemaPrompt("Extract the name: Alice", {
  name: "Person", schema: {
    type: "object", properties: { name: { type: "string" } },
    required: ["name"], additionalProperties: false
  }
});

schemaTransport accepts responses (default) or chat. If absent, decisionTransport is respected. Schema requests make one inference call and never silently fall back to JSON mode or another endpoint. Compatible services must support the selected schema wire format. Existing URL, Azure routing, authentication and ordinary prompt behavior continue to apply.

Responses uses text.format; Chat Completions uses response_format.json_schema. The explicit schema takes precedence over generic parameters; conflicting text, response_format, tools, input, messages and model parameters are rejected. noResponseFormat disables schema requests. Explicit nonempty tools are unsupported by this hook; registered tools are not invoked.

A private Ajv validator checks results without coercion, defaults or property removal. The service’s supported schema subset still applies. Refusals, incomplete output, unexpected tool calls and invalid results fail without committing conversation history. Statistics accompany successful results through jsonSchemaPromptWithStats.

Reference: https://developers.openai.com/api/docs/guides/structured-outputs