GuidesAI development

Tutorial: getting reliable JSON from an AI model

By the Banzai Labs teamUpdated on 2 min read

{ "nom": … "total": … "date": …}Else, asked againAn enforced schema, a checked output
Free text goes in, JSON matching the schema comes out, and the code validates it.

1. Describe the expected result

import { z } from 'zod';

const Invoice = z.object({
  supplier: z.string(),
  total: z.number(),
  date: z.string().regex(/^\d{4}-\d{2}-\d{2}$/),
});

2. Constrain the model

From loosest to strictest
  1. 1Instruction and exampleThe format described in the prompt, with an example.
  2. 2Forced toolA tool whose parameters follow the schema, which the model must call.
  3. 3Structured outputThe provider's mode that forces the output to follow the schema, where it exists.

3. Always validate

const parsed = Invoice.safeParse(JSON.parse(modelText));
if (!parsed.success) {
  // send parsed.error.issues back to the model and ask again, once or twice
} else {
  save(parsed.data);
}

The traps

  • Numbers written as text (“1,234.50 €”): state the expected format.
  • Fields missing from the document: allow null rather than forcing an invention.
  • Dates: enforce a single format.

See also hallucinations and tool use.

Frequently asked questions

Is writing “answer in JSON” in the prompt enough?
Often, not always. Structured output or forced-tool mechanisms make deviations much rarer. Validation in code is still mandatory.
What if validation fails?
Send the model the exact error and ask again, once or twice. Beyond that, set the case aside for a human.
Is valid JSON necessarily correct?
No. Well-formed JSON can hold a wrong value. Critical values are cross-checked against your data.

Read next

All guides