⚖️ LawDecoder

LawDecoder is an AI-powered legal explainer for Indian law, built using RAG (Retrieval-Augmented Generation). It retrieves relevant law sections (IPC, BNSS, IT Act, etc.) and explains them clearly, empathetically, and practically.

Backend runs on Node.js (Express) with local embeddings (via @xenova/transformers) + OpenRouter API for LLM inference. Frontend is a Streamlit app (Python) providing a simple, no-login interface.

---

🚀 Features

• ✅ Ask in plain English: Natural queries like "Someone built a house on my land".
• ✅ Grounded in law: Answers cite relevant sections (e.g., BNSS Section 167).
• ✅ Step-by-step guidance: Explains what to do (FIR, police, documents).
• ✅ Local embeddings: Runs on AMD GPU (WebGPU) or CPU fallback.
• ✅ Feedback system: Users can thumbs up/down and comment on responses.
• ✅ Privacy-first: Queries aren’t stored except optional feedback.

⚙️ Setup & Installation

1️⃣ Clone Repository

git clone https://github.com/ishwar170695/LawDecoder.git
cd LawDecoder

---

2️⃣ Backend (Node.js API)

Install dependencies:

cd backend
npm install

Configure environment:

Fill in your OpenRouter keys:

OPENROUTER_API_KEY1=your_first_key
OPENROUTER_API_KEY2=your_second_key   # (Optional, auto-rotates if multiple keys)

Start backend:

npm start

Backend runs at: http://localhost:8000

---

3️⃣ Frontend (Streamlit UI)

Install dependencies:

cd ../frontend
pip install -r requirements.txt

Run UI:

streamlit run app.py

UI will open in browser at: http://localhost:8501

---

🧠 How It Works (RAG Pipeline)

1. User Query → Embedding: Query is converted to a vector using local MiniLM model (Xenova/all-MiniLM-L6-v2).
2. Vector Search: Finds top k matching sections in parsed_laws_vectors.json.
3. LLM Call: Query + retrieved sections are passed to OpenRouter (Qwen, Gemini, DeepSeek).
4. Answer Generation: AI responds empathetically with cited sections + clear guidance.
5. Feedback Loop: Users can thumbs up/down, stored in feedback.json.

---

📦 Data & Embeddings

• Law texts (IPC, BNSS, IT Act, Constitution) are pre-parsed into JSON in backend/data/.

• Embeddings (parsed_laws_vectors.json) are too large for GitHub and must be generated locally:

  node backend\utils\cache_embedding.js
  node backend\utils\embed_parsed_laws.js

  (Script will use MiniLM for embeddings and save them locally.)

---

🔒 Privacy & Disclaimer

• No tracking: Queries aren’t logged beyond session.

• Optional feedback: Thumbs-up/down is stored locally in feedback.json.

• Disclaimer:

  LawDecoder is for informational purposes only and is not legal advice. Always consult a qualified lawyer for disputes.

---

🖥 Requirements

• Backend: Node.js 18+, npm, AMD GPU (optional, WebGPU) or CPU.
• Frontend: Python 3.8+, streamlit.
• LLM API: OpenRouter free-tier or personal keys.

---

🛠 Tech Stack

• Backend: Node.js, Express, @xenova/transformers (local embeddings)
• Frontend: Python, Streamlit
• LLMs: Qwen 72B, Gemini Flash, DeepSeek R1 (via OpenRouter)
• RAG: Vector search over parsed law JSON

---

🤝 Contributing

PRs welcome!

• Keep PRs focused (UI, backend, docs, etc.).
• Use .gitignore to avoid committing large files (parsed_laws_vectors.json).
• For major law updates (e.g., new Acts), run local embedding script.

---

📜 License

MIT License.