Practitioner open-sources Meridian document parsing pipeline, hits 118 pages/min on one H200

An enterprise OCR and analysis pipeline designed to process large multi-file job bundles (up to 50,000 pages) in regulated industries. LLMs are used for document analysis after OCR extraction. It addresses scaling and reliability problems when naive designs fail on high-volume, multi-PDF jobs.

Releases·Tool Release

PyMuPDF4LLM adds Image Analyzer turning PDF images into LLM-ready Markdown, from GRAG paper author

A new Image Analyzer module for PyMuPDF4LLM extracts charts, formulas, tables, and logos from PDFs and converts them into clean, LLM-ready Markdown, going beyond OCR by classifying image type and formatting accordingly. The module implements the image-analyzer approach from the author's GRAG research paper on group relative attention guidance for document parsing and RAG, and ships as an open-source Python package.

A new Image Analyzer module for PyMuPDF4LLM extracts charts, formulas, tables, and logos from PDFs and converts them into clean, LLM-ready Markdown, going beyond OCR by classifying image type and formatting accordingly. The module implements the image-analyzer approach from the author's GRAG research paper on group relative attention guidance for document parsing and RAG, and ships as an open-source Python package.

Products·launched

Local PDF-to-RAG parser for messy technical docs on 8GB VRAM

A two-stage local pipeline that converts messy PDFs (scans, nested tables) into markdown for RAG. It routes pages with clean text to a text extractor and renders complex pages for a vision model (Ollama, qwen3.5:9b). Runs on an 8 GB VRAM laptop, solving parsing issues for technical documentation.

A two-stage local pipeline that converts messy PDFs (scans, nested tables) into markdown for RAG. It routes pages with clean text to a text extractor and renders complex pages for a vision model (Ollama, qwen3.5:9b). Runs on an 8 GB VRAM laptop, solving parsing issues for technical documentation.

Products·building

DStudio: open-source desktop app routing tasks to specialized local models

DStudio is an open-source desktop app that uses DeepSeek V4 as the main reasoning model while routing specialized tasks to smaller local models: Qwen2.5-VL for image reading, Qwen Image for image generation/editing, Qwen3 Embedding for semantic document search, and Poppler for PDF text extraction. It solves the inefficiency of loading all capabilities into one large model, enabling semantic PDF retrieval for long documents.

DStudio is an open-source desktop app that uses DeepSeek V4 as the main reasoning model while routing specialized tasks to smaller local models: Qwen2.5-VL for image reading, Qwen Image for image generation/editing, Qwen3 Embedding for semantic document search, and Poppler for PDF text extraction. It solves the inefficiency of loading all capabilities into one large model, enabling semantic PDF retrieval for long documents.

DocSlicer is a lightweight, deterministic document parser and chunker for local agent/RAG pipelines. It converts PDF, DOCX, PPTX, and HTML into chunks, structured tables, and a heading hierarchy, processing 100-page files in about 2 seconds without models or GPU. It solves the parsing bottleneck for developers handling large financial and legal documents.