OpenDataLoader PDF

Feature Positioning

• Core Capability: Extract structured data (Markdown, JSON, HTML) from any PDF with bounding box coordinates
• Technical Highlights: XY-Cut++ reading order, Bounding Box positioning, hybrid AI mode for processing complex pages
• Benchmark Results: Overall score 0.90 (ranking 1st), table processing score 0.93, reading order score 0.94 (compared with Docling, Marker, MinerU, etc.)
• License: Apache 2.0 (core features are free)

Applicable Scenarios

• Batch extract PDF content to Markdown / JSON / HTML for RAG or LLM training
• Require bounding box coordinates for source tracing (identify which page and position of the PDF a paragraph comes from)
• Process complex tables, scanned documents, academic PDFs with formulas
• PDF accessibility (Tagged PDF generation, free to access in Q2 2026)

Installation

Prerequisites

• Java 11+
• Python 3.10+

pip install -U opendataloader-pdf

pip install "opendataloader-pdf[hybrid]"

Quick Start

CLI (suitable for single file or batch processing)

# 快速模式：输出 Markdown + JSON
opendataloader-pdf input.pdf output_dir/

# 指定格式
opendataloader-pdf input.pdf output_dir/ --format markdown,json,html

# 混合 AI 模式（复杂表格 / 扫描件）
opendataloader-pdf --hybrid docling-fast input.pdf output_dir/

# 混合模式 + OCR（扫描件）
opendataloader-pdf --hybrid docling-fast --force-ocr input.pdf output_dir/

# 混合模式 + 公式识别
opendataloader-pdf --hybrid docling-fast --hybrid-mode full input.pdf output_dir/

Python API

import opendataloader_pdf

# 批量处理（一次调用会启动 JVM，建议批量一次性传入）
opendataloader_pdf.convert(
    input_path=["file1.pdf", "file2.pdf", "folder/"],
    output_dir="output/",
    format="markdown,json"
)

Mode Selection Guide

opendataloader-pdf file.pdf out/

opendataloader-pdf --hybrid docling-fast file.pdf out/

--force-ocr

--force-ocr --ocr-lang "ko,en"

--hybrid docling-fast --hybrid-mode full

--enrich-picture-description --hybrid-mode full

Output Format Description

Markdown

JSON (with bounding boxes)

{
  "pages": [{
    "page_number": 1,
    "elements": [{
      "type": "heading",
      "text": "...",
      "bbox": [x0, y0, x1, y1],
      "level": 1
    }, {
      "type": "table",
      "bbox": [x0, y0, x1, y1],
      "html": "..."
    }]
  }]
}

bbox

HTML

Gotchas

• Each

  convert()

  call starts a new JVM process, so it is recommended to pass all batch files at once instead of calling it multiple times in a loop
• Hybrid mode requires starting the server in the background:

  opendataloader-pdf-hybrid --port 5002

  , then add

  --hybrid docling-fast

  to the client command
• For

  --enrich-formula

  or

  --enrich-picture-description

  , you must add

  --hybrid-mode full

  to both the hybrid server and client commands, otherwise the enhancement features will be skipped silently
• After modifying Java options, you must run

  npm run sync

  , which will regenerate

  options.json

  and all Python/Node.js bindings

Comparison with Other Tools

Reference Information

• PyPI：

  pip install opendataloader-pdf

• npm：

  npm install @opendataloader/pdf

• Maven：

  org.opendataloader:opendataloader-pdf-core

• GitHub：https://github.com/opendataloader-project/opendataloader-pdf
• Benchmark tests：https://github.com/opendataloader-project/opendataloader-bench