MinerU — Extract LLM-Ready Data from Any Document

MinerU Architecture & Features

Processing Pipeline

Input Document (PDF/Scan/Image)
    │
    ├─ Layout Detection (LayoutLMv3)
    │   └─ Identify: text, tables, figures, formulas, headers
    │
    ├─ OCR Engine (PaddleOCR / Tesseract)
    │   └─ Extract text from scanned pages
    │
    ├─ Table Recognition
    │   └─ Convert tables to Markdown/HTML/LaTeX
    │
    ├─ Formula Recognition (UniMERNet)
    │   └─ Convert math formulas to LaTeX
    │
    └─ Output Assembly
        ├─ Markdown (with images)
        ├─ JSON (structured blocks)
        └─ Content list (flat text)

Key Capabilities

Output Formats

Markdown — Clean, LLM-friendly format:

# Chapter 1: Introduction

The experiment results shown in **Table 1** demonstrate...

| Metric | Model A | Model B |
|--------|---------|---------|
| F1     | 0.92    | 0.87    |

The loss function is defined as:

$$L = -\sum_{i} y_i \log(p_i)$$

JSON — Structured blocks with metadata:

{
  "blocks": [
    {"type": "title", "text": "Chapter 1: Introduction", "level": 1},
    {"type": "text", "text": "The experiment results..."},
    {"type": "table", "cells": [["Metric", "Model A"], ["F1", "0.92"]]},
    {"type": "equation", "latex": "L = -\\sum_{i} y_i \\log(p_i)"}
  ]
}

CLI Commands

# Auto-detect PDF type (text vs scanned)
magic-pdf -p paper.pdf -o output/ -m auto

# Force OCR mode for scanned documents
magic-pdf -p scan.pdf -o output/ -m ocr

# Text-only mode (faster, no OCR)
magic-pdf -p textbook.pdf -o output/ -m txt

# Batch process a directory
magic-pdf -p papers/ -o output/ -m auto

Performance Benchmarks

Tested on academic papers, financial reports, and legal documents:

• Text extraction accuracy: 95%+ on clean PDFs
• Table recognition: 90%+ F1 on complex tables
• Processing speed: ~2-5 pages/second on GPU, ~0.5-1 page/second on CPU
• Language support: 80+ languages via PaddleOCR

FAQ

Q: What is MinerU? A: MinerU is an open-source document extraction tool with 57,900+ GitHub stars that converts PDFs and scanned documents into clean Markdown or JSON for LLM and RAG applications, with high-fidelity layout detection, table extraction, and formula recognition.

Q: How is MinerU different from Docling or Marker? A: MinerU focuses on layout-aware extraction with deep learning models (LayoutLMv3) and excels at complex multi-column academic papers. Docling (IBM) has broader format support. Marker is faster but less accurate on complex layouts. MinerU has the highest star count (57K+) and strongest community.

Q: Is MinerU free? A: Yes, open-source under AGPL-3.0. Free for personal and academic use. Commercial use requires compliance with AGPL terms or a commercial license.