The Gemini API can return JSON constrained to a schema, so you skip brittle string parsing. With the google-genai SDK you pass a response_format on the Interactions API and hand it your schema. The OpenAI equivalent is structured outputs; if you have not called Gemini yet, start with your first Gemini API request.
1. Install google-genai and pydantic
pip install -U google-genai pydantic
Pydantic gives you a schema and client-side validation in one model.
2. Define the schema as a Pydantic model
from pydantic import BaseModel
class Recipe(BaseModel):
name: str
minutes: int
ingredients: list[str]
A plain Pydantic model whose JSON Schema you send as the response contract.
3. Request JSON with response_format
from google import genai
client = genai.Client() # reads GEMINI_API_KEY
interaction = client.interactions.create(
model="gemini-3.8-flash",
input="Give me a recipe for pancakes.",
response_format={
"type": "text",
"mime_type": "application/json",
"schema": Recipe.model_json_schema(),
},
)
response_format sets the MIME type to JSON and pins the output to your schema.
4. Parse the response into the typed object
recipe = Recipe.model_validate_json(interaction.output_text)
print(recipe.name, recipe.minutes)
model_validate_json turns the JSON text into a typed Recipe you can use directly.
5. Constrain fields with enums
from typing import Literal
from pydantic import BaseModel
class Ticket(BaseModel):
priority: Literal["low", "medium", "high"]
hours: int
A Literal becomes a JSON Schema enum, so the model must pick one of the allowed values.
Gemini enforces enum, minimum, and maximum on numbers, and minItems and maxItems on arrays; constraints like minLength it does not enforce, so keep validating with Pydantic after parsing. Structured output combined with tools is preview and limited to the Gemini 3 series. Model ids change: take the current one from the structured output docs or the models page.
Verify it worked
interaction.output_text should be valid JSON, and model_validate_json should return a Recipe without raising.
If it fails:
- ValidationError from Pydantic — the model returned a field your schema forbids; loosen the model or add the field, then re-parse.
- Plain prose instead of JSON — you omitted
mime_type: "application/json"; both the MIME type and the schema are required to force JSON. - Unknown model error — the model id is stale; copy the current id from the Gemini models page.
Next: OpenAI structured outputs · glossary: structured output.