Force an LLM to return JSON in JavaScript

The safest way to get JSON from an LLM is to use a schema-based response format.

This example asks the model to classify a support message and return only the fields your app needs.

Install

npm install openai zod

Set your API key:

export OPENAI_API_KEY="your_api_key_here"

Code

import OpenAI from 'openai'
import { zodTextFormat } from 'openai/helpers/zod'
import { z } from 'zod'

const openai = new OpenAI()

const TicketLabel = z.object({
  category: z.enum(['billing', 'bug', 'feature_request', 'other']),
  priority: z.enum(['low', 'medium', 'high']),
  summary: z.string(),
})

const response = await openai.responses.parse({
  model: 'gpt-5.5',
  input: [
    {
      role: 'system',
      content: 'Classify the support ticket. Keep the summary under 20 words.',
    },
    {
      role: 'user',
      content: 'My invoice was charged twice this month. Please fix it quickly.',
    },
  ],
  text: {
    format: zodTextFormat(TicketLabel, 'ticket_label'),
  },
})

const ticket = response.output_parsed

console.log(ticket.category)
console.log(ticket.priority)
console.log(ticket.summary)

Example output:

{
  category: 'billing',
  priority: 'high',
  summary: 'Customer was charged twice and needs billing correction.'
}

Why not just JSON.parse?

You can ask a model to "return JSON", but plain prompting can still produce invalid JSON, extra text, or missing fields.

Schema-based output is better when the result goes into real application code.

Use it for:

• Ticket classification
• Lead enrichment
• Entity extraction
• Moderation labels
• App-ready LLM responses

Common mistakes

• Prompting "respond in JSON" and then calling JSON.parse on the raw text. It works in a demo, then breaks the first time the model wraps the JSON in prose or a code fence. The schema format above removes that whole class of bug.
• Forgetting the model can still refuse. A structured response can come back empty if the request is rejected, so check for that before you read fields.
• Making every field optional. If a field is required for your code to work, mark it required in the schema so a missing value fails loudly instead of silently becoming undefined.

Related

• OpenAI structured outputs with Zod in JavaScript: the same approach framed around typed extraction and reusing the schema as your app types.
• OpenAI function calling in JavaScript: when you want the model to call your code, not just return data.

References

• OpenAI structured outputs docs
• OpenAI text generation docs