PDF-Extract-Kit
Toolbox of models for layout detection, formula recognition, OCR, and table parsing of complex PDFs.
About
MinerU's document parsing quality rests on PDF-Extract-Kit, OpenDataLab's toolbox of state-of-the-art models for taking complex PDFs apart. It bundles layout detection with DocLayout-YOLO, YOLO-v10, and LayoutLMv3 to locate text, titles, images, tables, and formulas; inline and block formula detection plus UniMERNet for converting formula images to LaTeX; PaddleOCR for text extraction with position data; StructTable-InternVL2-1B for turning tables into LaTeX, HTML, or Markdown; and a reading order module. Each task is exposed as a module, so pipelines can mix and match components rather than adopt a monolithic converter, and Read the Docs covers installation, model weight downloads, and per-algorithm evaluation. Setup uses a conda environment with Python 3.10 and pip-installed requirements, with a separate CPU-only requirements file available, though a GPU is the standard configuration for real throughput. The license is AGPL-3.0, inherited from the YOLO code and PyMuPDF dependencies, and the repository has about 9,800 stars. Choose it for building custom extraction pipelines; choose MinerU for ready-made PDF to Markdown conversion.
Should you use PDF-Extract-Kit?
Pick it when
Use PDF-Extract-Kit when you are building a custom extraction pipeline and want to mix separate modules for layout detection, formula-to-LaTeX, OCR, and table parsing rather than adopt one monolithic converter.
Look elsewhere when
Skip it if you just want PDFs turned into Markdown; MinerU is the maintained ready-made tool, while this repo has had no commits since January 2025. Its AGPL-3.0 license also rules it out for many closed products.
Alternatives to PDF-Extract-Kit
- MinerU
Ready-made converter built on this toolkit and still actively developed, with far less assembly but less freedom to swap individual models.
- Docling
MIT-licensed modular converter with its own layout and table models, avoiding AGPL obligations in commercial products.
- PaddleOCR
Apache-2.0 option with layout, table, and formula modules in PP-StructureV3, and the OCR engine this kit already uses.
- Camelot
Light, MIT, CPU-friendly table extraction for digital PDFs when tables are all you need and a model stack is overkill.
Reviews (0)
Leave a Review
No reviews yet. Be the first to review!
Details
- Category
- OCR & Document Processing
- Price
- Free
- Platform
- Local/Desktop
- Difficulty
- Intermediate (3/5)
- License
- AGPL-3.0
- Added
- Jul 29, 2026