ExtractThinker

ExtractThinker is an open-source document intelligence library that applies large language models to structured data extraction from documents, combining traditional document parsing with LLM-powered field identification to produce structured outputs from unstructured document formats. Where traditional extraction tools use template-based rules that break when document layouts change, ExtractThinker uses LLMs to identify fields from context: specify what you want (invoice number, vendor name, line items, total amount) and the library finds it regardless of where it appears on the page or how the document is formatted. The library supports defining extraction schemas as Python dataclasses, making the output strongly typed and integration-ready without post-processing or JSON parsing. It handles PDFs, images, and Office documents through multiple vision and parsing backends.

The Splitter module processes multi-page documents intelligently, breaking them at semantic boundaries (each page, each document in a batch file) before extraction. ExtractThinker chains extraction steps with classification and splitting steps in configurable pipelines, handling common real-world document processing patterns like invoice batches and contract portfolios. Available as a Python package under the Apache License 2.0. Commonly used in accounts payable automation, contract analysis, insurance claims processing, and any workflow where structured data must be reliably extracted from variable-format business documents at scale.