MinerU

OpenDataLab’s open-source document parsing engine that converts complex documents — PDF, DOCX, PPTX, XLSX, images, web pages — into LLM-ready Markdown / JSON. Built for RAG · Agent · pre-training pipelines. Born inside the InternLM pre-training effort to solve symbol conversion in scientific literature.

Links

Description

High-accuracy parser with a VLM + OCR dual engine and 109-language OCR. Outputs human reading order, strips headers/footers/page numbers, reconstructs layout. Formulas → LaTeX, tables → HTML. Handles scanned docs, handwriting, multi-column layouts, and cross-page table merging.

Download or use

• Repo: https://github.com/opendatalab/mineru
• Online / desktop / API: https://mineru.net/
• Install: uv pip install -U "mineru[all]" (Python 3.10–3.13)
• Run: mineru -p <input> -o <output> (add -b pipeline for pure-CPU)
• Docs: https://opendatalab.github.io/MinerU/

Reasoning for

The document ingestion front-end for any knowledge-base or RAG workflow. Where Unlimited-OCR is a raw OCR model and CocoIndex / LightRAG handle indexing + retrieval, MinerU sits earlier in the chain: turning messy real-world documents into clean structured Markdown that downstream tools can index. Direct fit for feeding sources into LLM Knowledge Bases, OpenKB, or this vault’s ingest pipeline.

Key reasons to reach for it:

• MCP server → wire it straight into Claude Code, Cursor, Claude Desktop, Windsurf — agents can parse documents on demand.
• Native RAG-framework integration — LangChain, LlamaIndex, RAGFlow, RAG-Anything, Flowise, Dify, FastGPT.
• Fully offline / private — pipeline backend runs CPU-only (4GB VRAM or none), no hallucination. vlm/hybrid backends (vLLM/SGLang/LMDeploy/mlx) trade hardware for ~95 OmniDocBench accuracy.
• Permissive license — moved from AGPLv3 to the MinerU Open Source License (Apache-2.0 based) in v3.1, dropping commercial-adoption friction.

🧩 Backends

Backend	Best for	CPU	Accuracy (OmniDocBench v1.6)
pipeline	Fast, stable, no hallucination	✅	86.47
hybrid	High accuracy + native text extraction	❌	95.39 (high) / 95.26 (medium)
vlm	Max accuracy, vLLM/LMDeploy/mlx	❌	95.30

hybrid defaults to effort=medium (35–220% faster than high for ~0.13-point accuracy loss; high adds image analysis). Latest VLM model: MinerU2.5-Pro-2605-1.2B.

Alternatives considered

• Unlimited-OCR — Baidu 3B OCR model; lower-level, OCR-only vs MinerU’s full layout/format pipeline.
• Open Notebook — NotebookLM clone; ingests PDFs but for chat/podcast, not structured extraction.
• Commercial doc-AI APIs — MinerU trades polish for open-source + offline control.
• Sibling OpenDataLab tools: PDF-Extract-Kit, Magic-Doc, OmniDocBench (the benchmark MinerU scores against).

Resources

• Changelog: https://opendatalab.github.io/MinerU/reference/changelog/
• DeepWiki Q&A: https://deepwiki.com/opendatalab/MinerU
• Papers: MinerU2.5-Pro (arXiv 2604.04771), MinerU-Diffusion (2603.22458), MinerU 2.5 (2509.22186), original MinerU (2409.18839)
• Related vault notes: LLM Knowledge Bases, OpenKB, CocoIndex, LightRAG, AI Chatbots Architecture
• OpenMed — downstream complement: parse documents with MinerU, then extract clinical entities / de-identify PII on-device

---

Template: tool