MinerU Documentation — High-Accuracy Document Parsing for RAG and Agents
by OpenDataLab
Turn messy PDFs, scans and Office files into reading-order Markdown your RAG pipeline can actually use.
Overview
MinerU is an open-source document parsing engine from OpenDataLab that converts PDFs, scanned images, DOCX, PPTX and XLSX files into LLM-ready Markdown and JSON — the ingestion layer most open RAG stacks quietly skimp on. The documentation is organised as Quick Start (including extension modules and Docker deployment), Usage (quick usage, model source selection, CLI tools, advanced CLI parameters) and Reference (output file format, changelog), plus an FAQ and a hosted demo. The engine ships interchangeable backends: a CPU-friendly pipeline backend scoring 86.47 end-to-end on OmniDocBench v1.6, and hybrid/VLM backends reaching 95.26-95.39 by running a small vision-language model through vLLM or LMDeploy. Parsing preserves human reading order, strips headers, footers, footnotes and page numbers, converts formulas to LaTeX and tables to HTML, and handles scanned pages, handwriting and multi-column layouts with OCR across 109 languages. It runs on pure CPU or with CUDA, Ascend NPU or Apple MPS acceleration, and exposes a CLI, a FastAPI server and a Gradio web UI. Release 3.4 (18 June 2026) upgraded OCR to PP-OCRv6 for roughly 11% better accuracy at about double the speed; 3.3 added parameterised parsing strength worth 35-220% speedups depending on device. The project carries roughly 78,000 GitHub stars and ships under the MinerU Open Source License, based on Apache 2.0 with additional conditions.
At a Glance
- Topic
- RAG
- Level
- Intermediate
- Format
- Documentation
- Cost
- Free
- Duration
- ~2-3 hours to work through the docs and run a first parse; ongoing as reference
- Provider
- OpenDataLab
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Convert PDFs, images, DOCX, PPTX and XLSX into reading-order Markdown and JSON
- ✓Choose between the CPU pipeline backend and the higher-accuracy hybrid VLM backends
- ✓Serve MinerU as a FastAPI endpoint or Gradio web UI for a whole team
- ✓Drive the CLI with advanced parameters to trade parsing strength against throughput
- ✓Extract formulas as LaTeX and tables as HTML so downstream chunking keeps structure
- ✓Strip headers, footers, footnotes and page numbers before embedding any document
- ✓Deploy with Docker and configure model sources for offline or air-gapped installs
- ✓Read the output file format so your chunker can use layout and block metadata
Highlights
- •Two swappable backends let you trade measured OmniDocBench accuracy (86.47 vs ~95.3) against CPU cost
- •OCR spans 109 languages and copes with scanned, handwritten and multi-column source documents
- •Roughly 78,000 GitHub stars, with releases 3.3 and 3.4 landing within months of each other
- •Runs on plain CPU as well as CUDA, Ascend NPU and Apple MPS — no forced GPU spend to get started
- •Publishes its own OmniDocBench numbers per backend instead of vague accuracy marketing
Who It’s For
Best For
- ✓RAG engineers whose retrieval quality is limited by poor PDF and table extraction
- ✓Teams building document-ingestion services for regulated or air-gapped environments
- ✓Data engineers assembling LLM pretraining or fine-tuning corpora out of document archives
- ✓Agent developers who need a document-to-Markdown tool their agent can call
Prerequisites
- •Comfortable with Python 3.10+, pip and the command line
- •Basic understanding of RAG chunking and embedding pipelines
- •GPU access is optional — needed only for the hybrid and VLM backends
FAQ
What is MinerU Documentation — High-Accuracy Document Parsing for RAG and Agents?
MinerU is a production-grade, open-source document parsing engine that turns PDFs, scanned images and Office files into clean Markdown and JSON for RAG pipelines, agent tools and LLM pretraining corpora. It is aimed at engineers whose retrieval quality is bottlenecked by bad extraction rather than by embeddings or reranking. After working through the docs you can stand up a CPU or GPU parsing service, choose between the pipeline and VLM backends, and get reading-order-correct text with LaTeX formulas and HTML tables out of genuinely messy real-world documents.
Is MinerU Documentation — High-Accuracy Document Parsing for RAG and Agents free?
MinerU Documentation — High-Accuracy Document Parsing for RAG and Agents is free to access.
What level is MinerU Documentation — High-Accuracy Document Parsing for RAG and Agents for?
MinerU Documentation — High-Accuracy Document Parsing for RAG and Agents is aimed at a intermediate audience. Recommended background: Comfortable with Python 3.10+, pip and the command line, Basic understanding of RAG chunking and embedding pipelines, GPU access is optional — needed only for the hybrid and VLM backends.
How long does MinerU Documentation — High-Accuracy Document Parsing for RAG and Agents take?
Expect roughly ~2-3 hours to work through the docs and run a first parse; ongoing as reference. Most learners work through it at their own pace.
What will I learn from MinerU Documentation — High-Accuracy Document Parsing for RAG and Agents?
You'll learn: Convert PDFs, images, DOCX, PPTX and XLSX into reading-order Markdown and JSON; Choose between the CPU pipeline backend and the higher-accuracy hybrid VLM backends; Serve MinerU as a FastAPI endpoint or Gradio web UI for a whole team; Drive the CLI with advanced parameters to trade parsing strength against throughput; Extract formulas as LaTeX and tables as HTML so downstream chunking keeps structure; Strip headers, footers, footnotes and page numbers before embedding any document; Deploy with Docker and configure model sources for offline or air-gapped installs; Read the output file format so your chunker can use layout and block metadata.
Topics
Sources
This page was written from 3 sources, 2 on domains other than opendatalab.github.io.