PDF识别提取专家

OpenDataLoader PDF Integration — Reference#

Text-layer PDF parsing: JSON with bounding boxes, Markdown, HTML, Text, Tagged PDF. Supports table detection, reading order (XY-Cut++), sanitization.

Environment#

Requires Java 11+ and the opendataloader-pdf PyPI package (~21.5MB, includes 23MB Java CLI JAR).

java -version 2>&1 | head -1  # must be 11+
python3 -c "import opendataloader_pdf; print('✅')" 2>&1 || echo "⚠️ not installed"

Installation#

# 1. Java (if missing)
sudo apt-get install -y -qq default-jre

# 2. Python SDK (use Tsinghua mirror for speed — official PyPI often times out)
python3 -m venv ~/.venvs/opendataloader
~/.venvs/opendataloader/bin/pip install -i https://pypi.tuna.tsinghua.edu.cn/simple opendataloader-pdf

# 3. Optional: add to PATH
ln -sf ~/.venvs/opendataloader/bin/opendataloader-pdf ~/.local/bin/

Key Finding: Official PyPI (files.pythonhosted.org) downloads frequently time out; Tsinghua mirror delivers at 60+ MB/s.

CLI Usage#

# Basic: JSON + Markdown
opendataloader-pdf input.pdf -f json,markdown -o ./output_dir

# All formats
opendataloader-pdf input.pdf -f json,markdown,html,text,tagged-pdf -o ./output_dir

# Stdout (single format only)
opendataloader-pdf input.pdf -f markdown --to-stdout 2>/dev/null

# Page range
opendataloader-pdf input.pdf -f json -o ./output --pages "1,3,5-7"

# Sanitize (redact emails, phones, bank card numbers)
opendataloader-pdf input.pdf -f json -o ./output --sanitize

# Table detection mode
opendataloader-pdf input.pdf -f json -o ./output --table-method cluster

# Keep original line breaks
opendataloader-pdf input.pdf -f markdown -o ./output --keep-line-breaks

Python SDK#

from opendataloader_pdf import convert
convert(
    "/path/to/contract.pdf",
    output_dir="/path/to/output",
    format="json,markdown,html",
    reading_order="xycut",      # XY-Cut++ reading order (default)
    table_method="default",     # or "cluster"
    pages="1,3,5-7",
    sanitize=False,
    keep_line_breaks=False
)

Output Formats#

Format	Content	Use Case
json	type/hierarchy/content/bounding box/font/size/color per element	Programmatic processing, positional traceability
markdown	Structured text with headings, lists, tables	LLM input, quick reading
html	Layout-preserving page	Web preview
text	Plain text	Search, simple analysis
tagged-pdf	Tagged PDF	Accessibility (PDF/UA)

Performance#

• Local deterministic mode: ~0.015s/page (pure Java, no API calls)
• AI hybrid mode: Complex pages (OCR) need extra Hybrid service
• No network required: Local mode fully offline
• 23MB JAR: Large initial download

Limitations#

• Source PDF must have extractable text layer (scanned docs need PaddleOCR path — see main pdf-ocr-md skill)
• Scanned OCR recommended path: PaddleOCR (Chinese) or OpenDataLoader hybrid mode