The document excavator that feeds your RAG pipeline

What it does

MinerU is a document parsing engine that ingests PDFs, Word docs, PowerPoints, Excel sheets, images, and web pages, then emits structured Markdown or JSON tuned for LLM consumption. It handles the gnarly details—multi-column layouts, scanned pages, handwriting, cross-page tables, formulas rendered as LaTeX, and headers/footers stripped out—so downstream RAG and agent workflows get clean context instead of layout soup. The project ships with three inference backends: a fast CPU/GPU pipeline, a high-accuracy VLM engine, and a hybrid mode that tries to keep hallucinations low.

The interesting bit

The real signal is the dual VLM+OCR engine and the recent license shift from AGPLv3 to a custom Apache 2.0–based license, which lowers the barrier for commercial use. It also integrates as an MCP server for Cursor and Claude Desktop, and supports a laundry list of domestic Chinese AI chips, suggesting it was built for serious offline enterprise deployment, not just weekend hacking.

Key highlights

• Native parsing for DOCX, PPTX, and XLSX alongside PDFs and images; outputs Markdown, JSON, and HTML tables.
• VLM + OCR dual engine covering 109 languages, with recent upgrades for chart parsing and cross-page table merging.
• Runs fully offline with three backend options: fast pipeline, VLM-driven accuracy, or a hybrid low-hallucination mode.
• Plugs directly into LangChain, LlamaIndex, Dify, and FastGPT, plus an MCP server for AI coding tools.
• Supports 10+ domestic AI accelerators (Ascend, Cambricon, MetaX, etc.) in addition to standard NVIDIA/CPU stacks.

Caveats

• The custom “MinerU Open Source License” is Apache 2.0–based but not standard OSI Apache 2.0, so legal review is warranted before commercial embedding.
• The README is heavy on feature lists and light on quantitative accuracy benchmarks or latency numbers.

Verdict

Worth evaluating if you’re building RAG or agentic workflows that ingest messy real-world documents, especially in regulated or air-gapped environments. Skip it if your documents are already clean text or you prefer to outsource parsing to a managed API.