Instructor Library

Define your desired output structure using standard Pydantic models with Python type hints. Instructor maps these directly to the model's function calling or JSON mode capabilities.

• Uses BaseModel for schema definition
• Supports nested objects, enums, and optional fields
• Automatically generates JSON Schema for the LLM

When a model returns malformed JSON or data that fails Pydantic validation, Instructor catches the error and automatically re-prompts the model with the specific validation failure message.

• Configurable max_retries parameter
• Passes ValidationError context back to the LLM
• Dramatically increases extraction reliability

Instructor provides a unified patching interface across major LLM providers. The same Pydantic schema works whether you're calling OpenAI, Anthropic, Cohere, or open-source models via litellm.

• instructor.from_openai()
• instructor.from_anthropic()
• Consistent API regardless of backend

Extract structured data incrementally as tokens arrive. Instructor can yield partial Pydantic objects that are progressively built, enabling real-time UI updates during long generations.

• Use create_partial() for iterative extraction
• Each yielded object is a valid but incomplete model
• Ideal for chat interfaces and dashboards

Beyond validation retries, Instructor integrates with the Tenacity library for robust exponential backoff on API errors, rate limits, and transient failures.

• Handles HTTP 429 and 5xx errors gracefully
• Configurable retry strategies
• Production-grade resilience out of the box

LLMs often return numbers as strings or dates in inconsistent formats. Instructor automatically coerces types to match your Pydantic field definitions, handling common serialization mismatches.

• String-to-int coercion
• ISO 8601 date parsing
• Custom validators via Pydantic's @field_validator