ThePipe is an open-source document extraction library that uses vision-language models to turn messy, real-world documents into clean, structured data. A popular open-source project on GitHub with around 1,500 stars and active development in 2026, it targets the documents that break conventional parsers: scanned PDFs with no text layer, irregular tables, figures, equations, and mixed layouts.
Instead of relying on fragile text-layer parsing, ThePipe renders document pages and passes them to state-of-the-art vision-language models, returning clean Markdown, structured data that conforms to a Pydantic schema you define, or CSV output. It handles PDFs, Word documents, PowerPoint decks, spreadsheets, images, and scanned files, locating tables and figures on the page with bounding boxes and returning them as cropped images with provenance metadata. An agentic double-checking step, with an optional human-in-the-loop workflow, helps audit extractions for mission-critical data. The library produces per-page or semantic chunks sized for vector databases such as ChromaDB and RAG frameworks such as LlamaIndex, and it can convert output into OpenAI message format ready for any OpenAI-compatible language model. It is MIT-licensed and can be deployed on-premises, including air-gapped setups with local models.
Freemium. The open-source library is free to self-host under the MIT license with your own model keys. The hosted API and cloud platform are paid: Hobby at $25/month with 1M tokens, Scale at $210/month with 10M tokens, and custom Business plans with on-premise support.
Developers and data teams building document-processing or RAG pipelines who need accurate extraction from scanned, irregular, or layout-heavy documents that traditional parsers mangle.