AI Document Extractor Pro

A modern, PyQt6-powered desktop app for extracting, labeling, and exporting structured data from documents (PDFs, images) using advanced AI models (T5, LayoutLM, etc). Includes a no-code trainer and dataset builder for your own fine-tuning!

---

✨ Features

• Document Extraction:
  • Extracts key fields (invoice number, date, total, etc) from PDFs and images using LayoutLM and T5 models.
  • Supports both pre-trained and your own fine-tuned models (auto-discovered from models/).
• Beautiful UI:
  • Modern dark theme, responsive layout, and real-time logs.
• Export:
  • Save results as JSON, Excel, or SQLite with one click.
• Trainer GUI:
  • Fine-tune T5 models on your own data (no code needed).
  • Build datasets from PDFs/images with a simple labeling dialog.
• Post-processing:
  • Smart field correction and validation for cleaner results.

---

🚀 Quick Start

1. Clone the repo:

   git clone https://github.com/foudhilriahi/AIDocumentExtractorPro
   cd YOUR-REPO

2. Create a virtual environment:

   python -m venv myenv
   myenv\Scripts\activate  # On Windows
   # source myenv/bin/activate  # On Linux/Mac

3. Install dependencies:

   pip install -r requirements.txt

4. Run the app:

   python main.py

5. (Optional) Train your own model:

   python trainer.py

---

🧠 How It Works

• Extraction:
  • Select a document, choose a model, and click "Start Extraction".
  • See logs and results in the UI.
• Trainer:
  • Build a dataset from your own PDFs/images (label each sample in the dialog).
  • Train a new model and it will appear in the main app automatically.

---

📁 Project Structure

main.py                # Main app UI
trainer.py             # Trainer and dataset builder
ai_extractor.py        # Extraction logic
export_manager.py      # Export logic
requirements.txt       # Dependencies
models/                # Your models (auto-discovered)
logs/                  # Logs

---

📦 Dataset Format

• Each dataset is a folder containing a train.json and val.json (or just a single JSON for small sets).
• Each JSON is a list of samples like:

{
  "input": "text or OCR text here",
  "output": "structured JSON string here"
}

You can adapt this to your real dataset as needed.

---

🏷️ Supported Fields (for extraction/training)

• INVOICE_RECEIPT_ID
• INVOICE_RECEIPT_DATE
• DUE_DATE
• TOTAL
• TAX
• SUBTOTAL
• VENDOR_NAME
• VENDOR_ADDRESS
• SHIP_TO_NAME
• SHIP_TO_ADDRESS
• BILL_TO_NAME
• BILL_TO_ADDRESS

---

✅ Capabilities Checklist

• Display clean AI extraction results from flan-t5-base, flan-t5-small, and layoutlm-invoices
• Save to .json, .xlsx, .db
• Show key:value results in a rich readable text viewer
• Handle file uploads and GPU/CPU logs smartly
• Modern UI
• No ugly logs, clean messages only
• Future-ready for training, doc type detection, etc.

---

🏋️‍♂️ Training Workflow

• Provides a real TrainingThread class
• Trains LayoutLMv3 or FLAN-style models using Hugging Face’s transformers
• Automatically processes FUNSD/SROIE-style annotated datasets
• Emits progress and final result signals
• Future-ready for GUI integration (e.g. phase 3 training window in app_gui.py)

To train:

• Select your dataset folder (with FUNSD/SROIE-style JSON + image, or your own format)
• Choose model: LayoutLMv3 vs custom T5 (bonus FLAN training later)
• Training runs in a background thread and updates the GUI

---

🛠️ Requirements

• Python 3.9+
• PyQt6
• transformers, datasets, evaluate, torch, easyocr, pytesseract, pdf2image, pillow

---

🤝 Contributing

Pull requests and issues are welcome! Please open an issue for bugs or feature requests.

---

📜 License

MIT

---

🙏 Credits

• Hugging Face Transformers
• PyQt6
• LayoutLM, T5, and all open-source model authors

---

Made with ❤️ by [FOUDHIL_RIAHI]