instructor

To extract structured JSON from LLM outputs, define a Pydantic response model with your required types and constraints, then pass it as the response_model parameter. The LLM output is parsed and validated against your schema, returning a fully typed object.

LLM structured data extraction fails validation when outputs miss fields or violate constraints. This is fixed by enabling automatic retries, which feed the specific Pydantic validation errors back to the LLM as context, prompting regeneration until the output conforms.

Yes, you can stream partial JSON objects from an LLM during generation. By enabling streaming, you receive partially constructed Pydantic objects incrementally as the LLM produces tokens, allowing you to power responsive UIs or incremental data pipelines.

Yes, this approach works with both OpenAI and Anthropic APIs for structured output. It provides consistent Pydantic validation and structured extraction patterns across multiple supported LLM providers, ensuring type safety regardless of the underlying API.

The best way to parse nested JSON structures and complex types from natural language is using Pydantic schemas with automatic retries. This safely handles nested objects, optional fields, and strict constraints, ensuring downstream code can trust the validated result.

To enforce type safety when extracting multiple entities from LLM responses, define a Pydantic model containing nested objects, enums, and constraints. The extraction process validates the LLM output against this schema, automatically retrying with error feedback if validation fails.