ragtable-extract 0.1.0

PDF Parsing for RAG — extracting tables precisely. Convert to HTML. Fast, local, no GPU.

ragtable-extract

English | 中文

PDF Parsing for RAG — extracting tables precisely. Convert to HTML. Fast, local, no GPU.

A lightweight Python library that extracts tables from PDFs and converts them to clean HTML, designed for RAG pipelines and LLM retrieval. Runs entirely on CPU with no external APIs or GPU dependencies.

Features

• Precise extraction — Character-level coordinate extraction for accurate cell boundaries
• ≥/≤ symbols — Correctly repositioned by coordinates (avoids pdfplumber's line-end placement bug)
• Merged cells — Proper rowspan / colspan output for multi-column tables
• Line-wrapped text — Auto-segments and concatenates text across line breaks within cells (no symbol/text serialization)
• Fangzheng font — Handles full-width character ordering and decimal point encoding issues
• Adaptive config — Per-page tuning based on character metrics
• Fast & local — Pure Python, pdfplumber-based, no GPU required

Requirements

• Python 3.8+
• pdfplumber >= 0.10.0

Installation

pip install ragtable-extract

Or from source:

git clone https://github.com/ZhuJiaxin2/ragtable-extract.git
cd ragtable-extract
pip install -e .

Quick Start

import ragtable_extract

# Convert PDF tables to HTML file
ragtable_extract.convert(
    input_path="document.pdf",
    output_path="tables.html",
)

# Or extract as structured data
tables = ragtable_extract.extract(input_path="document.pdf")
for t in tables:
    print(f"Page {t['page']}: {t['html'][:80]}...")

CLI

python -m ragtable_extract document.pdf output.html

Web Quick Test (app.py)

Run the Flask web app to upload PDFs and preview extraction results in the browser:

pip install flask
python app.py

Then open http://localhost:1965 to upload a PDF and view extracted tables.

Test Results

Run python test.py to generate extraction results. Output files:

Source PDF	Extraction Result
test/example/zhejiang.pdf	test/result/test_adaptive_zhejiang.html
test/example/changsha.pdf	test/result/test_adaptive_changsha.html
test/example/shaanxi.pdf	test/result/test_adaptive_shaanxi.html
test/example/tongbao.pdf	test/result/test_adaptive_tongbao.html

API

Function	Description
convert(input_path, output_path, pages?, config?, use_adaptive_config=True)	Convert PDF tables to HTML file
extract(input_path, pages?, config?, use_adaptive_config=True)	Extract tables as list of dicts with page, html, bbox, raw
build_full_html(pdf_filename, tables)	Build full HTML document from extracted tables
Config	Dataclass for tuning extraction (multiline thresholds, font tolerance, etc.)

Configuration

import ragtable_extract

# Custom config
config = ragtable_extract.Config(
    multiline_cell_top_range=25,
    multiline_y_tolerance=4,
)
tables = ragtable_extract.extract("doc.pdf", config=config)

# Adaptive config (default) — infers parameters from page character metrics
tables = ragtable_extract.extract("doc.pdf")  # use_adaptive_config=True by default

Project Structure

ragtable-extract/
├── ragtable_extract/     # Core library
│   ├── __init__.py       # convert(), extract()
│   ├── _core.py          # Table extraction logic
│   ├── _config.py        # Config & adaptive metrics
│   ├── _font.py          # Special font handling
│   └── _html.py          # HTML template
├── pyproject.toml
├── demo.py               # CLI demo
└── app.py                # Optional Flask web API

How It Works

PDF → pdfplumber.find_tables()
  → Filter chars by bbox, cluster by top (y)
  → Reorder ≥/≤ symbols, fix Fangzheng font
  → Output <table> HTML

Comparison Report

We compare ragtable-extract with opendataloader-pdf on real government PDF tables. Our extraction:

• Multi-column tables — Correctly recognizes complex layouts with merged cells
• Line-wrapped text — Automatically segments and concatenates text across line breaks within cells
• No serialization — Symbols and text stay in correct cells (e.g. no １ ２ or 万人 ％ wrongly merged)

Run python test_comparison.py to generate the report, then open comparison_report.html for side-by-side results.

License

MIT