Instructor

Instructor is the standard Python library for structured LLM outputs: you define a Pydantic model describing exactly the data you want, and Instructor coerces any LLM's response into a validated, typed object. Built around the unified instructor.from_provider() interface, the same code works across OpenAI, Anthropic, Google, Mistral, Cohere, Ollama, DeepSeek, xAI, Cerebras, Fireworks, and more providers, with automatic retries when validation fails, streaming of partial responses, and full IDE type inference. It is a local client-side library (your API keys stay in your environment), so it adds no network or telemetry surface of its own. The project was formerly jxnl/instructor and moved to the 567-labs org (Jason Liu's company); the old repo path still redirects.

1. Install: pip install instructor (or uv add instructor / poetry add instructor). For a specific provider SDK, use an extra, e.g. pip install "instructor[xai]".
2. Export your provider API key, e.g. export OPENAI_API_KEY=... (or the key your provider expects).
3. Define a Pydantic model for the data you want, e.g. class Person(BaseModel): name: str; age: int.
4. Build a client: client = instructor.from_provider("openai/gpt-5-nano") (swap the provider string for others, e.g. "anthropic/claude-3", "ollama/llama3.2"; pass async_client=True for async).
5. Extract: person = client.create(response_model=Person, messages=[{"role": "user", "content": "Extract: John is a 30-year-old software engineer"}]) — returns a validated Person instance.