Instructor

Getting typed data out of a chat model is the whole job here. You declare a Pydantic class, pass it as response_model, and the library handles schema translation, the provider call, parsing and validation; when a field fails a validator, the error text goes back to the model and the request is retried up to a limit you choose. One from_provider() entry point covers OpenAI, Anthropic, Gemini, Vertex AI, Mistral, Cohere, Groq, xAI, Bedrock and several more, plus local models through Ollama or any OpenAI-compatible server. It can stream partially filled objects as tokens arrive, copes with nested models and lists, and ships a small CLI for batch jobs, file uploads and usage. Setup is pip install instructor on Python 3.9 or newer, with pydantic 2 and the openai SDK as core dependencies; no GPU is involved because inference runs wherever your provider does. MIT licensed, so commercial use is unrestricted, and ports exist for TypeScript, Ruby, Go, Elixir and Rust. Best for developers who need dependable extraction or classification without adopting a full agent framework; the maintainers point to PydanticAI for agent-heavy work.