promptdojo_

Schemas, Pydantic, and validation — making the model return real data — step 3 of 9

promptdojo_›phase 03 · llm apis›ch 14 · structured output
lesson 1 of 2 · schemas, pydantic, and validation — making the model return real datastep 3 / 9

Pydantic models — the schema is the contract

A Pydantic model is just a Python class where every attribute has a type hint. That class is the contract between your prompt and your downstream code:

from pydantic import BaseModel

class Ticket(BaseModel):
    email: str
    severity: int
    summary: str

Three things this gives you that AI relies on:

  1. JSON schema for free. Ticket.model_json_schema() returns the exact schema you can pass to the model in a tool definition or response_format. The model now knows what shape you expect.
  2. One-line parse + validate. Ticket.model_validate_json(raw) parses the JSON string, type-checks every field, raises ValidationError if anything is off. No manual isinstance checks.
  3. Real Python objects. After validation you write ticket.email, not ticket["email"]. Your IDE autocompletes the field names and catches typos at edit time.

Three field-type idioms you'll see constantly

from typing import Literal
from pydantic import BaseModel, Field

class Issue(BaseModel):
    title: str
    severity: Literal["low", "medium", "high"]   # enum
    score: int = Field(ge=0, le=10)              # bounded int
    tags: list[str] = []                         # default empty list

Literal[...] for enums (the model is forced to pick one of three strings). Field(ge=..., le=...) for numeric bounds. Default values make a field optional. AI uses all three on every real schema.

Run the editor. We mimic Pydantic with a plain dict + assertions — identical logic, just lower-level — so you can see what model_validate_json does for you in one call.

read, then continue.