कानून (Kanoon) is an AI-powered legal assistant chatbot designed to help users with Indian legal queries. It leverages a Retrieval-Augmented Generation (RAG) pipeline combined with a Large Language Model (LLM) to provide accurate, context-aware legal information based on the Indian Constitution, legal cases.
- Real-time Legal Q&A - Get instant answers to legal questions
- Hybrid RAG/LLM Mode - Automatically switches between database-backed and general knowledge responses
| Technology | Purpose | Version |
|---|---|---|
| React | UI Framework | 18.3.1 |
| TypeScript | Type Safety | 5.8.3 |
| Vite | Build Tool & Dev Server | 5.4.19 |
| Shadcn UI | Component Library | Latest |
| Tailwind CSS | Styling | 3.4.17 |
| React Router | Navigation | 6.30.1 |
| Lucide React | Icons | 0.462.0 |
| Technology | Purpose | Version |
|---|---|---|
| FastAPI | Web Framework | Latest |
| Python | Programming Language | 3.12+ |
| Uvicorn | ASGI Server | Latest |
| ChromaDB | Vector Database | Latest |
| Sentence Transformers | Embeddings | Latest |
| Ollama | Local LLM Runtime | Latest |
| Qwen 3 | Language Model | 8B parameters |
| Technology | Purpose |
|---|---|
| Hugging Face Datasets | Data Loading |
| Beautiful Soup | Web Scraping (optional) |
| Pandas | Data Manipulation |
| tqdm | Progress Bars |
Ith/
├── backend/
│ ├── models/
│ │ ├── main.py # FastAPI application & API endpoints
│ │ ├── rag_pipeline.py # RAG pipeline implementation
│ │ ├── vector_database.py # ChromaDB vector database manager
│ │ ├── data/
│ │ │ └── vectordb/ # ChromaDB persistent storage
│ │ │ ├── chroma.sqlite3
│ │ │ └── e1567276.../ # Vector embeddings
│ │ └── __pycache__/
│ │
│ ├── scrapers/
│ │ ├── dt.py # Hugging Face dataset loader
│ │ └── data/
│ │ └── raw/
│ │ └── legal_data_all.json # Raw legal Q&A dataset
│ │
│ ├── requirements.txt # Python dependencies
│ └── .env # Environment variables
│
├── frontend/
│ ├── src/
│ │ ├── components/
│ │ │ ├── ChatInterface.tsx # Main chat component
│ │ │ ├── Navigation.tsx # Header navigation
│ │ │ ├── Footer.tsx # Footer component
│ │ │ └── ui/ # Shadcn UI components
│ │ │
│ │ ├── services/
│ │ │ └── api.ts # API service for backend communication
│ │ │
│ │ ├── pages/
│ │ │ ├── Index.tsx # Landing page
│ │ │ └── NotFound.tsx # 404 page
│ │ │
│ │ ├── lib/
│ │ │ └── utils.ts # Utility functions
│ │ │
│ │ ├── App.tsx # Main app component
│ │ └── main.tsx # App entry point
│ │
│ ├── package.json # Node dependencies
│ ├── vite.config.ts # Vite configuration
│ ├── tailwind.config.ts # Tailwind configuration
│ └── .env # Frontend environment variables
│
└── DOCUMENTATION.md # This file
┌─────────────┐ HTTP/REST ┌──────────────┐
│ Frontend │ ◄─────────────────► │ FastAPI │
│ (React) │ JSON Data │ Backend │
└─────────────┘ └──────────────┘
│
│
▼
┌─────────────────────────────────────┐
│ RAG Pipeline Manager │
└─────────────────────────────────────┘
│ │
┌─────────┴──────────┴─────────┐
│ │
▼ ▼
┌─────────────┐ ┌─────────────┐
│ ChromaDB │ │ Ollama │
│ Vector DB │ │ (Qwen3) │
└─────────────┘ └─────────────┘
Embeddings Search Text Generation
- API Service: Handles all HTTP requests to backend
- State Management: React hooks (useState, useEffect) for local state
- CORS Configuration: Allows frontend-backend communication
- Request Validation: Pydantic models for type safety
- Embedding Model:
sentence-transformers/all-MiniLM-L6-v2 - Search: Cosine similarity search for relevant Q&A pairs
- Metadata: Stores question, answer, and other metadata
- Document Retrieval: Fetches top-k similar Q&A pairs from ChromaDB
- Hybrid Mode: Switches between RAG and pure LLM based on similarity threshold
- LLM Generation: Sends prompt to Ollama (Qwen3) for response generation
- Local Execution: Runs on user's machine (no API costs)
- Model: Qwen3 8B - optimized for chat and instruction following
- Temperature: 0.2 for more factual, consistent responses
- Max Tokens: 1500 tokens (~1000-1200 words)
Retrieval-Augmented Generation is a technique that combines:
- Information Retrieval - Finding relevant documents from a database
- Text Generation - Using an LLM to generate responses based on retrieved context
graph TD
A[User Query] --> B[Generate Query Embedding]
B --> C[Search Vector DB]
C --> D{Similarity > Threshold?}
D -->|Yes| E[RAG Mode]
D -->|No| F[Pure LLM Mode]
E --> G[Retrieve Top-K Q&A Pairs]
G --> H[Build Context Prompt]
H --> I[Send to Ollama LLM]
F --> J[Build General Prompt]
J --> I
I --> K[Generate Response]
K --> L[Return to User with Sources]
# User asks: "What are fundamental rights?"
query = "What are fundamental rights in Indian Constitution?"# Convert query to vector representation
query_embedding = embedding_model.encode(query)
# Result: [0.123, -0.456, 0.789, ...] (384 dimensions)# Search ChromaDB for similar Q&A pairs
results = vector_db.search(query_embedding, top_k=5)
# Returns: Top 5 most similar questions with their answers# Calculate average similarity
avg_similarity = 1 - avg_distance
threshold = 0.35 # Configurable
if avg_similarity >= threshold:
mode = "RAG" # Use database context
else:
mode = "LLM" # Use general knowledgeRAG Mode (with context):
System: You are an AI legal assistant...
Context from Database:
[Reference 1] (ID: qa_123):
Question: What are fundamental rights?
Answer: Fundamental rights are basic human rights enshrined in Part III...
**LLM Mode (no context):**
System: You are an AI legal assistant with expertise in Indian law... Note: This response is based on general knowledge.
User: What are fundamental rights in Indian Constitution? Assistant:
#### Step 6: Response Generation
```python
# Send prompt to Ollama
response = ollama.generate(
model="qwen3:8b",
prompt=prompt,
temperature=0.2,
max_tokens=1500
)
┌──────────────────────────────────────────────────────────────────┐
│ USER INTERACTION │
└───────────────────────────┬──────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ 1. USER TYPES QUERY │
│ Example: "What is Article 21 of Indian Constitution?" │
└───────────────────────────┬─────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ 2. FRONTEND SENDS REQUEST │
│ POST /api/chat │
│ Body: { query, conversation_id?, top_k? } │
└───────────────────────────┬─────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ 3. BACKEND RECEIVES REQUEST │
│ - Validates input (max 1000 chars) │
│ - Checks RAG pipeline status │
└───────────────────────────┬─────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ 4. GENERATE QUERY EMBEDDING │
│ - Use Sentence Transformer model │
│ - Convert text to 384-dim vector │
└───────────────────────────┬─────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ 5. SEARCH VECTOR DATABASE (ChromaDB) │
│ - Cosine similarity search │
│ - Retrieve top-5 Q&A pairs │
│ - Calculate similarity scores │
└───────────────────────────┬─────────────────────────────────────┘
│
▼
┌───────────┴───────────┐
│ │
▼ ▼
┌───────────────────┐ ┌───────────────────┐
│ Similarity >= 0.35│ │ Similarity < 0.35 │
│ RAG MODE │ │ LLM MODE │
└─────────┬─────────┘ └─────────┬─────────┘
│ │
▼ ▼
┌─────────────────┐ ┌─────────────────────┐
│ Build RAG Prompt│ │ Build General Prompt│
│ - Add Q&A pairs │ │ - No database info │
│ - Add sources │ │ - General knowledge │
└─────────┬───────┘ └─────────┬───────────┘
│ │
└──────────┬──────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ 6. SEND PROMPT TO OLLAMA (Qwen3:8b) │
│ - Temperature: 0.2 │
│ - Max tokens: 1500 │
│ - Timeout: 120s │
└───────────────────────────┬─────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ 7. OLLAMA GENERATES RESPONSE │
│ - Processes prompt │
│ - Generates coherent answer │
│ - Returns text response │
└───────────────────────────┬─────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ 8. FORMAT RESPONSE WITH SOURCES │
│ - Extract source metadata │
│ - Format citations │
│ - Update conversation history │
└───────────────────────────┬─────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ 9. SEND RESPONSE TO FRONTEND │
│ { │
│ response: "Article 21 states...", │
│ sources: [...], │
│ conversation_id: "uuid", │
│ timestamp: "2025-10-29T..." │
│ } │
└───────────────────────────┬─────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ 10. FRONTEND DISPLAYS RESPONSE │
│ - Shows bot message with answer │
│ - Displays source cards below │
│ - Maintains conversation context │
└─────────────────────────────────────────────────────────────────┘
- Python 3.12+
- Node.js 18+
- Ollama installed
- Git
git clone <repository-url>
cd Ith# Navigate to backend
cd backend
# Create virtual environment
python -m venv venv
source venv/bin/activate # On Windows: venv\Scripts\activate
# Install dependencies
pip install -r requirements.txt
# Set up environment variables
# Create .env file in backend directory with:
# FRONTEND_ORIGINS=http://localhost:5173
# OLLAMA_MODEL=qwen3:8b
# OLLAMA_BASE_URL=http://localhost:11434
# VECTOR_DB_PATH=./data/vectordb# Navigate to scrapers
cd scrapers
# Run dataset loader
python dt.py
# This will download the dataset from Hugging Face and save it
# Follow the prompts to process all examples# Navigate to models
cd ../models
# Run vector database setup
python vector_database.py
# This will:
# 1. Create ChromaDB instance
# 2. Generate embeddings for all Q&A pairs
# 3. Store in persistent database# Pull the Qwen3 8B model
ollama pull qwen3:8b
# Verify it's running
ollama list# In backend/models directory
python -m uvicorn main:app --reload --port 8000
# Or simply
python main.py# Navigate to frontend
cd ../../frontend
# Install dependencies
npm install
# Create .env file with:
# VITE_API_BASE_URL=http://localhost:8000
# Start development server
npm run devOpen browser and navigate to:
- Frontend:
http://localhost:5173 - Backend API Docs:
http://localhost:8000/docs
- User authentication and sessions
- Chat history persistence
- Export chat as PDF
- Multi-language support
- Voice input/output
- Radhakrishna Bharuka -
- Nilesh Dwivedi-
- Hari Krishna Sharma-
This project is licensed under the MIT License.
- Hugging Face - For the
viber1/indian-law-dataset
For questions or support, please reach out:
- Email: radhebharuka29@gmail.com