Introduction

Every retrieval-augmented generation pipeline eventually runs into the same obstacle: the PDF. Before a language model can extract answers from contracts, research papers, or invoices, something must convert images, layouts, and scanned pages into machine-readable text. A new ranking of 138 open source OCR and PDF parsing projects, each with more than 1,000 GitHub stars, offers developers a practical starting point for choosing the right tool.

The list comes from a directory that monitors roughly 3,141 high-starred open source text mining repositories. After filtering for optical character recognition and document parsing, the curators narrowed the field to 138 active projects. Star totals reflect a snapshot taken on November 11, 2026.

What Happened

PaddleOCR leads the ranking with approximately 92,600 stars, edging out MinerU at roughly 83,500 and Tesseract at about 77,500. The full set spans classic engines, browser-based tools, desktop screen grabbers, handwriting recognizers, and math-to-LaTeX converters.

The methodology excluded projects that only appeared to be OCR-related. From an initial pool of 321 repositories tagged for OCR and documents, the team removed 23 entries such as generic AI courses, inference frameworks, clipboard utilities, and captcha solvers. The remaining 138 all cleared the 1,000-star threshold.

Python dominates the ecosystem, powering 158 of the 321 tagged repositories, while C++ and TypeScript account for 21 apiece. Licensing is mostly permissive, with 111 MIT-licensed projects and 80 under Apache-2.0. However, notable exceptions include PyMuPDF under AGPL-3.0 and paperless-ngx under GPL-3.0, so teams should verify license compatibility before shipping.

A distinct trend is the shift toward PDF to LLM-ready Markdown. Projects such as MinerU, liteparse, zerox, and Dolphin position their output as structured data for AI pipelines rather than plain text extraction. Meanwhile, translation software for manga, visual novels, games, and subtitles represents a surprisingly large share of the top 50.

Vision models are also getting smaller. Xiaomi-OCR-0 handles text, tables, and formulas with 0.8 billion parameters, while TeleOCR performs structured document parsing with 1.2 billion parameters, pointing toward lighter on-device options.

Why This Matters

Document ingestion remains a bottleneck for AI applications. Traditional template-based OCR often fails when page layouts change, whereas vision-language models read documents more flexibly but can become expensive at scale. Many practitioners now favor a hybrid strategy that combines deterministic OCR with model-based extraction.

The ranking is useful beyond RAG pipelines. Developers can find browser-based OCR through tesseract.js and Paddle.js, make scanned PDFs searchable with OCRmyPDF, or build self-hosted paperless offices using paperless-ngx and docspell. Specialized tools cover table extraction, handwriting recognition, and equation conversion.

License choice carries real consequences. Permissive options simplify commercial integration, while copyleft licenses may require derivative works to remain open. Checking the license before deployment is essential for production systems.

Key Takeaways

  • PaddleOCR is the most-starred open source OCR project on the list, followed by MinerU and Tesseract.
  • Python is the leading language for OCR and document parsing repositories.
  • Most projects use MIT or Apache-2.0 licenses, but copyleft options such as AGPL-3.0 and GPL-3.0 appear among popular tools.
  • Newer projects increasingly output Markdown or JSON optimized for large language models.
  • Translation tools for comics, games, and subtitles make up a notable portion of the highest-ranked repositories.
  • Compact vision models like Xiaomi-OCR-0 and TeleOCR suggest efficient local OCR is becoming more viable.

Conclusion

Open source OCR has matured into a diverse ecosystem with options for nearly every document workflow. Whether the goal is searchable PDFs, structured data for AI agents, or offline desktop capture, the 138-project ranking provides a data-driven shortlist. Developers should still review individual licenses, confirm recent maintenance, and test candidates against their own document samples before committing to a production stack.

The full directory includes additional niche tools in the 100 to 999 star range, along with filters for programming language, license, and last update date.