Skip to main content

Structured output

Ask any provider for JSON and get a parsed object back. With a JSON Schema the provider's native structured output is used, so the reply matches the schema.

chatJson​

const { Chatbot, ChatGPTInput } = require('intellinode');

const schema = {
  type: 'object',
  properties: {
    city: { type: 'string' },
    country: { type: 'string' },
    population: { type: 'integer' },
  },
  required: ['city', 'country', 'population'],
};

const bot = new Chatbot(OPENAI_API_KEY, 'openai');
const input = new ChatGPTInput('Answer as JSON.', { responseSchema: schema });
input.addUserMessage('Where is the Eiffel Tower?');

const data = await bot.chatJson(input);
// { city: 'Paris', country: 'France', population: 2102650 }

chatJson sends the request, then parses the first reply. The reply is parsed even when a model wraps it in prose or code fences, and the result is validated against the schema type (an object schema rejects an array reply).

Input options​

  • responseSchema: a JSON Schema for the reply. Sent as native structured output on OpenAI (chat completions and the Responses API), Anthropic, Gemini, Mistral, Cohere and the OpenAI-compatible services.
  • responseFormat: 'json': free-form JSON without a schema (OpenAI json_object mode; on Claude it becomes a system instruction).
  • strictSchema: true: OpenAI strict mode. Every object gets additionalProperties: false and optional properties become required and nullable, as strict mode demands.

The same options work on AnthropicInput, GeminiInput, MistralInput, CohereInput, NvidiaInput, VLLMInput and OpenAICompatibleInput:

const { AnthropicInput, SupportedChatModels } = require('intellinode');

const claude = new Chatbot(ANTHROPIC_API_KEY, SupportedChatModels.ANTHROPIC);
const input = new AnthropicInput('Answer as JSON.', { responseSchema: schema });
input.addUserMessage('Where is the Colosseum?');

const data = await claude.chatJson(input);   // { city: 'Rome', country: 'Italy', population: ... }

Provider notes:

  • Gemini receives the schema without $schema and additionalProperties, which its dialect rejects.
  • Anthropic requires closed objects; intellinode adds additionalProperties: false where it is missing.
  • DeepSeek only supports json_object; the schema is added to the system message and the reply is still parsed and checked.

One-line JSON with Gen​

Gen.generate_json does the same in one call:

const { Gen } = require('intellinode');

const data = await Gen.generate_json('Where is the Colosseum?', schema, OPENAI_API_KEY, 'openai');

// free-form JSON without a schema
const facts = await Gen.generate_json('Three facts about Rome as {"facts": [...]}', null, ANTHROPIC_API_KEY, 'anthropic');

Parsing helpers​

OutputParser exposes the parsers used internally, useful when you call chat() directly:

const { OutputParser } = require('intellinode');

const object = OutputParser.parseJson('Sure! ```json\n{"a": 1}\n```', 'object');   // { a: 1 }
const code = OutputParser.extractCode('```js\nconsole.log(1)\n```', 'js');         // 'console.log(1)'