Instructor

Instructor is a Python library that patches LLM client libraries to return structured, validated outputs instead of raw text. Built on Pydantic, it lets you define a response model as a Pydantic class and get back a validated Python object — with automatic retries when the LLM output doesn't match the schema. It's not an agent framework; it's a precision tool for one specific problem: getting reliable structured data from LLMs.

The library works by patching the OpenAI, Anthropic, Google, Cohere, Mistral, and other client libraries with a responsemodel parameter. When you call client.chat.completions.create(responsemodel=MyModel, ...), Instructor handles the function-calling schema generation, response parsing, validation, and retry logic. If the LLM returns invalid data, Instructor feeds the validation errors back to the model and retries.

Instructor supports multiple extraction modes: TOOLS (native function calling), JSON (forces JSON output), MD_JSON (extracts JSON from markdown blocks), and PARALLEL (extracts multiple objects). TOOLS mode is most reliable with capable models, while JSON mode works better with models that have weak function calling.

Beyond basic extraction, Instructor supports streaming partial objects (incremental Pydantic model updates as the LLM generates), iterable responses (extract lists of objects), union types for classification, and validators with custom logic. The library also includes a citation validator for grounding extracted data.

Created by Jason Liu, Instructor has become the de facto standard for structured extraction in Python LLM applications, with ports to TypeScript, Ruby, Go, and Elixir.

The honest take: Instructor does one thing exceptionally well. If your challenge is getting structured, validated data from LLMs — entity extraction, classification, data transformation — Instructor is the right choice over heavier frameworks. It's the tool you reach for when you need a Pydantic model back from an LLM call, reliably.