---
title: Saudi Law AI Assistant
emoji: ⚖️
colorFrom: blue
colorTo: purple
sdk: docker
pinned: false
---

# Law Document RAG Chat Application
...

# Law Document RAG Chat Application

A web application that allows users to ask questions about indexed legal documents using Retrieval Augmented Generation (RAG) techniques.

## Features

- 🤖 **RAG-powered Q&A**: Ask questions about your legal documents and get answers extracted directly from the context
- 📚 **Document Indexing**: Automatically index PDF, TXT, DOCX, and DOC files from a folder
- 🎨 **Modern React Frontend**: Beautiful, responsive chat interface
- ⚡ **FastAPI Backend**: High-performance API with LangChain and FAISS
- 🔍 **Exact Context Extraction**: Answers are extracted directly from documents, not generated
- 🔀 **Hybrid Search**: Combines BM25 (keyword-based) and semantic search for improved retrieval accuracy
- 🤗 **Qwen Model Support**: Uses Qwen/Qwen3-32B model via HuggingFace router for high-quality Arabic language understanding
- 🚀 **Hugging Face Spaces Ready**: Configured for easy deployment

## Tech Stack

- **Frontend**: React 18
- **Backend**: FastAPI
- **RAG**: LangChain + FAISS with Hybrid Search (BM25 + Semantic)
- **Vector Database**: FAISS
- **Embeddings**: Qwen/Qwen3-Embedding-8B (HuggingFace) or OpenAI embeddings (configurable)
- **LLM**: Qwen/Qwen3-32B via HuggingFace router (default) or OpenAI API (configurable)
- **Hybrid Search**: BM25 + Semantic search using EnsembleRetriever
- **Python**: 3.10 or 3.11 (required for faiss-cpu compatibility)

## Project Structure

```
KSAlaw-document-agent/
├── backend/
│   ├── main.py              # FastAPI application
│   ├── rag_system.py        # RAG implementation
│   ├── document_processor.py # Document processing logic
│   ├── embeddings.py        # OpenAI embeddings wrapper
│   └── chat_history.py     # Chat history management
├── frontend/
│   ├── src/
│   │   ├── App.js           # Main React component
│   │   ├── App.css          # Styles
│   │   ├── index.js         # React entry point
│   │   └── index.css        # Global styles
│   ├── build/               # Built React app (for deployment)
│   ├── public/
│   │   └── index.html       # HTML template
│   └── package.json         # Node dependencies
├── documents/               # Place your PDF documents here
├── vectorstore/            # FAISS vectorstore (auto-generated)
├── app.py                   # Hugging Face Spaces entry point
├── Dockerfile               # Docker configuration
├── pyproject.toml           # Python dependencies (uv)
├── uv.lock                  # Locked dependencies
├── processed_documents.json # Processed document summaries
└── README.md                # This file
```

## Quick Start

**Local Development:**
1. Install dependencies: `uv sync` and `cd frontend && npm install`
2. Create `.env` file in the project root with required environment variables:
   - `HF_TOKEN`: Your HuggingFace API token (required for Qwen model)
   - `OPENAI_API_KEY`: Your OpenAI API key (required for document processing)
   - Optionally set `USE_HYBRID_SEARCH=true` to enable hybrid search (BM25 + Semantic)
3. Add documents to `documents/` folder
4. Run backend: `uv run python backend/main.py`
5. Run frontend: `cd frontend && npm start`

**Deployment to Hugging Face Spaces:**
1. Build frontend: `cd frontend && npm run build`
2. Set up Xet storage (recommended) or prepare to upload PDFs via UI
3. Push to Hugging Face: `git push hf main`
4. Set required environment variables in Space secrets:
   - `HF_TOKEN`: Your HuggingFace API token
   - `OPENAI_API_KEY`: Your OpenAI API key
   - Optionally set `USE_HYBRID_SEARCH=true` to enable hybrid search

## API Endpoints

- `GET /api/` - Health check
- `GET /api/health` - Health status
- `POST /api/index` - Index documents from a folder
  ```json
  {
    "folder_path": "documents"
  }
  ```
- `POST /api/ask` - Ask a question
  ```json
  {
    "question": "What is the law about X?",
    "use_history": true,
    "context_mode": "chunks",
    "model_provider": "qwen"
  }
  ```
  - `question` (required): The question to ask
  - `use_history` (optional): Whether to use chat history (default: `true`)
  - `context_mode` (optional): Context mode - `"full"` (entire document) or `"chunks"` (top semantic chunks, default)
  - `model_provider` (optional): Model provider - `"qwen"` (default) or `"openai"`
  
  **Note**: The default `model_provider` is `"qwen"` which uses Qwen/Qwen3-32B via HuggingFace router. When using `context_mode="chunks"` with hybrid search enabled, the system combines BM25 and semantic search for improved retrieval accuracy.

## Environment Variables

### Required Variables

- `HF_TOKEN`: Your HuggingFace API token (required for Qwen model and HuggingFace embeddings)
- `OPENAI_API_KEY`: Your OpenAI API key (required for document processing and optional for embeddings/LLM)

### Optional Configuration

- `QWEN_MODEL`: Qwen model to use (default: `Qwen/Qwen3-32B:nscale`)
- `EMBEDDINGS_PROVIDER`: Embeddings provider - `"openai"` or `"hf"`/`"huggingface"` (default: `"openai"`)
- `HF_EMBEDDING_MODEL`: HuggingFace embedding model (default: `Qwen/Qwen3-Embedding-8B`)
- `OPENAI_LLM_MODEL`: OpenAI LLM model to use (default: `gpt-4o-mini`)
- `OPENAI_EMBEDDING_MODEL`: OpenAI embedding model (default: `text-embedding-ada-002`)
- `USE_HYBRID_SEARCH`: Enable hybrid search combining BM25 and semantic search (default: `"false"`, set to `"true"` to enable)
- `HYBRID_BM25_WEIGHT`: Weight for BM25 component in hybrid search (default: `0.5`)
- `HYBRID_SEMANTIC_WEIGHT`: Weight for semantic component in hybrid search (default: `0.5`)
- `CHAT_HISTORY_TURNS`: Number of conversation turns to keep in history (default: `10`)

## Notes

- The system extracts exact text from documents, not generated responses
- Supported document formats: PDF, TXT, DOCX, DOC
- The vectorstore is saved locally and persists between sessions
- Documents are automatically processed on startup (no manual indexing needed)
- **Default Model**: The system uses Qwen/Qwen3-32B via HuggingFace router by default for better Arabic language understanding
- **Hybrid Search**: When enabled (`USE_HYBRID_SEARCH=true`), combines BM25 keyword search with semantic search for improved retrieval accuracy
- For Hugging Face Spaces, the frontend automatically uses `/api` as the API URL
- This project uses `uv` for Python package management - dependencies are defined in `pyproject.toml` and `uv.lock`
- The `.env` file should be in the project root (not in the backend folder)
- PDFs can be stored using Hugging Face Xet storage or uploaded via the Space UI

## Troubleshooting

### Common Issues

- **HF_TOKEN Error**: Make sure `HF_TOKEN` is set in your `.env` file (local) or Space secrets (deployment) when using Qwen model
- **OpenAI API Key Error**: Make sure `OPENAI_API_KEY` is set in your `.env` file (local) or Space secrets (deployment) for document processing
- **No documents found**: Ensure documents are in the `documents/` folder with supported extensions (PDF, TXT, DOCX, DOC)
- **Frontend can't connect**: Check that the backend is running on port 8000
- **Build fails on Spaces**: Ensure `frontend/build/` exists (run `npm run build`), check Dockerfile, verify dependencies in `pyproject.toml`
- **RAG system not initialized**: Check Space logs, ensure `processed_documents.json` exists and is not ignored by `.dockerignore`
- **Hybrid search not working**: Ensure `rank-bm25` is installed (`uv sync` should handle this) and `USE_HYBRID_SEARCH=true` is set

## License

MIT