Artifex Software

We've been doing more on benchmarking for PyMuPDF, read all about it here: https://lnkd.in/e5RdShEQ

We recently took a dip dive into benchmarking document parsing solutions and found some very interesting results: https://lnkd.in/dsn4FMUJ

With 10 years of PyMuPDF and over 1 billion downloads we asked the creator Harald Lieder for his personal perspective on this milestone. https://lnkd.in/eS9GAfwz

Our PyMuPDF package for document parsing with Python hit the 1 billion mark yesterday on PyPI !!! 🎉 https://lnkd.in/eZCj6TJf

Brotli is here! For the past few years, Artifex has worked with the PDF Association to bring Brotli compression to the PDF format, and the proposed spec is about to be published. What that means in practice: smaller PDFs at the same quality. The Adobe PDF 1.7 reference manual, 1,310 pages and 30.95MB, comes down to just over 13MB. A monochrome vector-and-text document recompresses to half its original size. Many popular PDF viewers have already been updated to cope with Brotli compression, including Ghostscript and MuPDF. Learn more in the full blog post: https://lnkd.in/eqJVZVyQ #PDF #Compression #MuPDF #Ghostscript #Brotli

With PyMuPDF4LLM 1.28.2, we move to a fully AGPL v3 licensing with the PyMuPDF Layout package dependency and all the source goes public. The fastest CPU-only layout engine for PDF is now genuinely open source. Read more: https://lnkd.in/g68Rd4qd

Tables remain one of the hardest problems in document intelligence. The latest release of PyMuPDF4LLM tackles this head-on with a major upgrade to table extraction. What's new: - Improved table bounding box detection - Full table structure detection - row segmentation, column segmentation, and cell grid reconstruction - Unified processing, so table understanding is now part of core layout analysis Read more: https://lnkd.in/gdAkefJt

PyMuPDF 1.28 is out, and it introduces a coordinated set of improvements that significantly expand its document-processing capabilities. Markdown is now a first-class document format, and Office files convert directly to Markdown. That means new workflows across every downstream tool, from PDF generation to document analysis. This blog post walks through the Markdown-to-PDF workflow with real CSS examples: https://lnkd.in/gNajRFza

A practical pattern for document pipelines: triage before you spend. PyMuPDF can extract per-page signals, like text density, image coverage, tables, form widgets, or layout complexity, in milliseconds, before anything reaches OCR or an LLM. You use those signals to classify each page (skip it, extract text directly, send to OCR, or send to an LLM) so the expensive tools only see the pages that justify them. Code + walkthrough in the full blog post: https://lnkd.in/gJWJZRTp

If your document AI system can tell a user "this clause is on page 4" but can't highlight the actual words on the page, you're missing the layer that makes the output trustworthy. The fix is coordinate-anchored extraction: storing bounding boxes alongside text and using them to verify the model's claims map to the expected page region. Learn more about the concept (grounding) and implementation with PyMuPDF: https://lnkd.in/geEEKNuT #LLM #RAG #DocumentAI #PyMuPDF