Use Zod or Pydantic to define structured output

Cost

Open source. Both are MIT licensed and free to use in any app.

How to do it

Zod is MIT licensed. It shipped 1.0 in March 2020 and 4.6.7 on 12 October 2026, with about 207 million weekly npm downloads. Pydantic is MIT licensed. It shipped 1.0 in October 2019 and 2.14.1 on 11 October 2026, with about 243 million weekly PyPI downloads. Both pass all five tests. Zod's z.toJSONSchema() and Pydantic's model_json_schema() both write JSON Schema draft 2020-12, which the provider SDKs accept. What to ask the model for is in Ask for output that matches a schema. Choose something else when the project uses another language. Then write the schema as a JSON Schema file and check replies with the most downloaded validator for that language.

Examples

Using these examples: they show one way to follow this rule. Keep the behaviour and the check, not the syntax. Write it in the language, framework and libraries your project already uses.

TypeScript example

import * as z from 'zod';

export const Ticket = z.object({
  category: z.enum(['billing', 'delivery', 'other']),
  summary: z.string().max(200),
});

export const ticketJsonSchema = z.toJSONSchema(Ticket);
export const parseTicket = (reply: string) => Ticket.parse(JSON.parse(reply));

Python example

from typing import Literal

from pydantic import BaseModel, Field

class Ticket(BaseModel):
    category: Literal["billing", "delivery", "other"]
    summary: str = Field(max_length=200)

TICKET_JSON_SCHEMA = Ticket.model_json_schema()

def parse_ticket(reply: str) -> Ticket:
    return Ticket.model_validate_json(reply)

How to check

Every model call whose answer code reads passes a schema. A unit test feeds the validator one valid reply and one invalid reply, and only the valid one passes.

Sources

• JSON Schema draft 2020-12

Related rules

• Ask for output that matches a schema in AI features