---
title: DocAI
emoji: 📊
colorFrom: blue
colorTo: gray
sdk: gradio
sdk_version: 6.6.0
app_file: app.py
pinned: false
license: apache-2.0
---

# 📊 DocAI - Document Intelligence Demo

A Hugging Face Space showcasing document parsing and chart intelligence using **Docling**, **Granite Vision**, and **Chart2CSV**.

## 🎯 Features

- **📄 Document Parsing** - Extract text & figures from PDFs using Docling
- **🖼️ Figure Detection** - Automatically identify and extract figures from documents
- **💬 Chart Q&A** - Ask questions about charts using Granite Vision
- **📊 Chart-to-CSV** - Convert charts to structured data (CSV tables)

## 🔄 Pipeline

```
Upload PDF
    ↓
Parse with Docling (extract text + figures)
    ↓
Select a figure/chart
    ↓
Ask questions OR Extract to CSV
```

## 🚀 Getting Started

### Local Testing

```bash
# Install dependencies
pip install -r requirements.txt

# Run app
python app.py
```

Then open http://localhost:7860

### Deploy to HF Spaces

```bash
# Commit and push (already set up as HF Space)
git add -A
git commit -m "Update DocAI demo"
git push
```

## 📦 Architecture

```
DocAI/
├── app.py                    # Main Gradio interface
├── requirements.txt          # Dependencies
├── src/
│   ├── ui_state.py          # State management & caching
│   ├── pdf_io.py            # PDF rendering (PyMuPDF)
│   ├── docling_parse.py     # Document parsing
│   ├── crops.py             # Figure extraction
│   ├── infer_vision_qa.py   # Granite Vision Q&A
│   ├── infer_chart2csv.py   # Chart2CSV extraction
│   └── utils.py             # Utilities
└── README.md
```

## 🤖 Models Used

- **Docling**: `ibm-granite/granite-docling-258M` - Document understanding
- **Granite Vision**: `ibm-granite/granite-vision-3.3-2b` - Image Q&A
- **Chart2CSV**: `ibm-granite/granite-vision-3.3-2b-chart2csv-preview` - Table extraction

## ⚙️ Configuration

### HuggingFace Hub

The space configuration is defined in `README.md` (YAML frontmatter):

- **SDK**: Gradio 6.6.0
- **App File**: `app.py`
- **License**: Apache 2.0

### Environment

No secrets required. All models are public on Hugging Face Hub.

## 📝 Usage Examples

### Chart Q&A
"What is the trend in this chart?" → Granite Vision analyzes and responds

### Chart2CSV
Upload a bar chart → Get structured CSV data with axis labels and values

## ⚠️ Known Limitations

- PDF parsing depends on Docling model accuracy
- Figure detection may miss complex multi-panel layouts
- Q&A quality depends on image clarity and model capabilities
- Large PDFs (100+ pages) may be slow on CPU

## 📄 License

Apache 2.0 - See LICENSE file

## 🙏 Credits

Built with:
- [Docling](https://github.com/DS4SD/docling) - IBM
- [Granite](https://huggingface.co/ibm-granite) - IBM AI Research
- [Gradio](https://gradio.app/) - Hugging Face