gurki309/contract-analysis-ai

contract-ai/
├── app/
│   ├── api/
│   │   ├── routes.py          # GET /health · POST /predict · POST /train
│   │   └── schema.py          # Pydantic request/response models
│   ├── core/
│   │   ├── config.py          # Pydantic Settings (env vars / .env)
│   │   └── model_loader.py    # Singleton model registry (loaded once)
│   ├── services/
│   │   ├── inference.py       # predict_clause(text) → label + confidence
│   │   ├── preprocessing.py   # Dataset load, clean, tokenize
│   │   └── training.py        # HuggingFace Trainer pipeline
│   └── main.py                # FastAPI app + lifespan startup
├── models/
│   ├── contract_classifier/   # Fine-tuned model weights (save here)
│   └── label_map.json         # { "0": "Termination", "1": "Indemnification", … }
├── data/                      # Raw / cached dataset files
├── notebooks/
│   └── colab_training.ipynb   # 🔥 Google Colab training notebook (GPU)
├── pyproject.toml
└── .env.example

git clone <your-repo-url>
cd legal-contract/contract-ai
cp .env.example .env

uv sync          # installs all dependencies from pyproject.toml

# Start server (model will be unavailable until training finishes)
uv run uvicorn app.main:app --host 0.0.0.0 --port 8000 --reload

# In another terminal — launch training
curl -X POST http://localhost:8000/api/v1/train \
     -H "Content-Type: application/json" \
     -d '{"num_train_epochs": 3, "per_device_train_batch_size": 16, "learning_rate": 2e-5}'

uv run python -m app.services.training

uv run uvicorn app.main:app --host 0.0.0.0 --port 8000 --reload

{
  "status": "ok",
  "model_loaded": true,
  "version": "1.0.0"
}

{ "text": "This agreement may be terminated by either party upon 30 days written notice." }

{ "label": "Termination", "confidence": 0.9821 }

{
  "num_train_epochs": 3,
  "per_device_train_batch_size": 16,
  "learning_rate": 2e-5,
  "test_size": 0.15
}

{
  "status": "training_started",
  "message": "Fine-tuning started in the background. Check GET /api/v1/health — when model_loaded=true the new model is ready."
}