Open Source Other Structured OutputPydanticPythonValidationJSON Mode

Instructor

Python library that patches any LLM client (OpenAI, Anthropic, Gemini) to return validated Pydantic models instead of raw JSON strings, with automatic retry on failure.

jxnl/instructor 10.5k stars 870 forks MIT

Install

$pip install instructor

About Getting Started Features References

About

Instructor wraps the client.chat.completions.create() call with a response_model parameter. Pass any Pydantic class and Instructor handles telling the model the schema, parsing the response, validating it, and retrying with the error message if validation fails.

It works with tool calling (for OpenAI/Anthropic) or JSON mode depending on the provider, always returning a proper Python object you can use immediately.

Getting Started

pip install instructor anthropic

import instructor
import anthropic
from pydantic import BaseModel

client = instructor.from_anthropic(anthropic.Anthropic())

class User(BaseModel):
    name: str
    age: int

user = client.messages.create(
    model="claude-sonnet-4-5",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Extract: John Doe is 29 years old"}],
    response_model=User,
)
print(user)  # User(name='John Doe', age=29)