RagTable

Extract tables from images/PDFs to Markdown — fast, accurate, RAG-ready

RagTable converts table images into structured Markdown using a segmentation model and per-cell OCR. Built for borderless and complex tables (academic papers, reports, scanned documents), with output that plugs directly into RAG pipelines or LLM contexts.

---

Features

• Table structure segmentation via EfficientUNet (~19M params) — detects row, col, col_header, row_header, and span masks
• Span detection — merged cells are identified and rendered correctly in Markdown
• Per-cell OCR via PaddleOCR — each cell is cropped and read individually to minimize noise
• Smart header handling — header rows get a dedicated OCR pass with spatial mapping for better accuracy
• Automatic post-processing — removes phantom edge columns, empty rows, and footer artifacts
• Fully offline — no external API calls, suitable for sensitive or air-gapped environments
• Multi-format input: PNG, JPG, TIFF

---

Installation

CPU (default)

pip install ragtab

This installs everything needed: PyTorch, PaddleOCR, PaddlePaddle (CPU), and other dependencies.

GPU (NVIDIA CUDA)

pip install ragtab[gpu]

This replaces paddlepaddle (CPU) with paddlepaddle-gpu for ~3-5× faster OCR inference on NVIDIA GPUs.

Requirements: Python ≥ 3.10

Install from source

git clone https://github.com/tai03102004/rag-table
cd ragtab
pip install -e .

---

Quickstart

from ragtab.pipeline import extract_table

markdown, cells = extract_table(
    "table.png",
    model_path="checkpoints/unet_best.pt",
    ocr_engine="paddleocr"
)

print(markdown)

Output:

| Item        | Price | Qty |
| ----------- | ----- | --- |
| iPhone 15   | 999   | 12  |
| Samsung S24 | 899   | 8   |

---

How It Works

Input image (resized to 384×384)
       │
       ▼
[1] EfficientUNet → 5 segmentation masks
       │
       ▼
[2] Projection analysis → row/column separator positions
       │
       ▼
[3] Span detection → connected components on span mask
       │
       ▼
[4] Grid construction → per-cell bounding boxes
       │
       ▼
[5] OCR — header rows: spatial mapping pass
        — body cells: per-cell crop + PaddleOCR
       │
       ▼
[6] Post-processing → drop phantom columns, empty rows, footers
       │
       ▼
[7] Markdown export

Each stage is independently accessible so you can customize or swap components.

---

Custom checkpoint

If you've trained your own model or want to use a different checkpoint:

markdown, cells = extract_table("table.png", model_path="path/to/your_model.pt")

---

Manual download (optional)

• Hugging Face (used by auto-download)
• Google Drive (mirror)

---

Cache location

By default, the model is cached at ~/.cache/ragtab/. Override via environment variable:

export RAGTAB_CACHE_DIR=/path/to/custom/cache

---

Project Structure

RagTable/
├── python/
│   └── ragtab/
│       ├── __init__.py
│       ├── detection.py     # Mask → grid cells
│       ├── model.py         # EfficientUNet definition
│       ├── ocr.py           # PaddleOCR wrapper + text cleaning
│       ├── pipeline.py      # End-to-end extract_table()
│       └── utils.py
├── checkpoints/
├── notebooks/
│   └── 02_table-recognition.ipynb
└── README.md

---

License

MIT — free to use, including for commercial purposes.

---

Author

Dinh Duc Tai — dinhductai2004@gmail.com

If you find this useful, consider giving it a ⭐️ on GitHub!