Skip to content
agentscamp
Tool

Instructor

Get structured, validated output from LLMs using plain type definitions, with automatic retries on validation failure.

open sourcesdk
Updated Jun 3, 2026
structured-outputpydanticvalidationopen-sourcepython
Visit website Source
View as Markdown

Instructor turns an LLM into a typed function: define a Pydantic model (or a Zod/equivalent schema in its ports), and Instructor coerces the model's output into that shape, validating it and automatically re-asking on failure. The simplest way to get reliable structured data out of an LLM.

Instructor makes structured output from LLMs feel like calling a typed function. You define the shape you want as a Pydantic model (Python — with ports for TypeScript, Go, and others), pass it in, and Instructor handles the rest: it instructs the model, parses the response into your type, validates it, and automatically retries with the validation errors fed back if the output doesn't conform.

It is aimed at developers who want data, not prose, from an LLM — extraction, classification, form-filling — without writing brittle JSON parsing and retry loops by hand. Because it builds on the providers' native function-calling/structured-output capabilities, it's thin and reliable rather than a heavy framework.

Highlights

  • Types as the schema — define output with Pydantic (or the language port's equivalent); no hand-written JSON Schema.
  • Validation + auto-retry — invalid output is re-requested with the errors, so you get conforming data or a clear failure.
  • Provider-agnostic — works across OpenAI, Anthropic, and many other models.
  • Streaming partials — stream structured objects as they're produced.
  • Minimal footprint — a focused library, not a framework you build your app around.

In an AI-assisted workflow

import instructor
from pydantic import BaseModel
 
class User(BaseModel):
    name: str
    age: int
 
client = instructor.from_provider("anthropic/claude")
user = client.chat.completions.create(response_model=User, messages=[...])  # -> validated User

TIP

Let your types do the prompting: a well-named model with field descriptions and constraints (enums, ranges) often beats paragraphs of instructions for getting the exact structure you want.

Good to know

Instructor is free and open source (MIT); you pay your model provider for tokens. For a cross-language, schema-first approach with its own DSL, compare BAML; for a TypeScript-native app toolkit that also does structured output, see the Vercel AI SDK. Background on the techniques: Structured Output vs JSON Mode vs Function Calling.

Related