⚖️ Agentic Legal AI (Confidence-Gated HITL)

A production-grade Sovereign AI assistant designed to provide expert-level legal advisory on the Constitution of India and updated Bharatiya Nyaya Sanhita (BNS) codes.

---

📺 Live Demo

🚀 Access the Expert Advisor

---

🏗️ Technical Architecture

The system is built on a modular Monolith-over-Microservices architecture, leveraging LangGraph for deterministic state-controlled RAG orchestration.

---

🚀 Key Engineering Features

🧠 Robust RAG Orchestration

• LangGraph State Machine: Deterministic flow control using a 4-node state graph (CLASSIFY → RETRIEVE → GENERATE → POST_PROCESS).
• Parent-Child Chunking: Implemented recursive character splitting with Jina AI embeddings. Child chunks (small) optimize vector search retrieval, while parent chunks (contextual) provide the LLM with surrounding knowledge to mitigate hallucinations.
• Confidence-Gated HITL Fallback: Dynamic thresholding (80% confidence cutoff). Automatically triggers an interactive Human-in-the-Loop Web Search via Tavily when the Vector DB similarity score is too low, effectively eliminating hallucination for out-of-database queries.
• Dual Search Strategy: Hybrid retrieval combining semantic vector search with keyword-based expansion for specific legal terminology (e.g., FIR, BNS Sections).

🛡️ Privacy & Reliability

• Intent Classification Guardrails: LangGraph nodes autonomously identify Prompt Injections, Time-Sensitive questions, and Out-of-Scope (OOS) queries before they reach the LLM, redirecting them to secure fallbacks.
• PII Masking Integration: Leverages Microsoft Presidio + spaCy to detect and anonymize names, phone numbers, and identifying data before transmission to the LLM.
• Circuit Breaker Pattern: Integrated pybreaker around external LLM calls (fail_max=10) to ensure system stability during upstream provider outages.
• Rate Limiting: IP-based rate limiting (5 req/min) via SlowAPI to prevent API abuse and control operational costs.

🔄 Intelligent Sync Engine

• Automated Re-indexing: Custom engine using SHA-256 hash comparison between Supabase Storage and a PostgreSQL registry.
• Orphan Cleanup: Automatic detection and deletion of vector embeddings for files removed from the knowledge base.
• Admin Dashboard: Real-time monitoring of document status (Indexed, Pending, Deleted) with UI-driven sync triggers.

⚡ Performance & Observability

• Three-Tier Caching: Response caching, active user tracking, and stream simulation via Upstash Redis.
• Observability: Real-time tracing of every user intent and LLM inference chain using Langfuse.
• GDPR Compliance: MongoDB-backed chat history with automatic 30-day TTL index for data-at-rest protection.

---

🛠️ Tech Stack

• Frontend: React 18, Tailwind CSS (Glassmorphism UI), Lucide Icons
• Backend: FastAPI, Uvicorn, LangChain, LangGraph, Pydantic
• AI Models: Qwen 3 235B (via OpenRouter), Jina AI Embeddings v2, Tavily Search API
• Infrastructure: Docker, Render (Compute), Qdrant Cloud (Vector Store), MongoDB Atlas (NoSQL), Supabase (Postgres + Object Storage), Upstash (Redis)

---

📂 Enterprise Folder Structure (Backend)

The backend follows a Domain-Driven Design (DDD) architecture to ensure scalability, security, and separation of concerns.

backend/
├── app/
│   ├── auth/          # Authentication, JWT, and Google OAuth
│   │   ├── __init__.py
│   │   ├── jwt.py
│   │   ├── oauth.py
│   │   ├── routes.py
│   │   └── schemas.py # Auth-specific Pydantic models
│   ├── core/          # Application Brain (Logging, Config, Rate Limiting)
│   │   ├── __init__.py
│   │   ├── config.py
│   │   ├── limiter.py
│   │   └── logger.py
│   ├── db/            # Database Clients (MongoDB, Supabase)
│   │   ├── __init__.py
│   │   ├── database.py
│   │   └── supabase_client.py
│   ├── rag/           # Retrieval-Augmented Generation Logic
│   │   ├── __init__.py
│   │   ├── graph.py   # LangGraph State Machine
│   │   ├── pipeline.py# Processing, Qdrant, Presidio PII
│   │   ├── routes.py  # Endpoints for RAG Chat
│   │   └── schemas.py # RAG-specific Pydantic models
│   ├── utils/         # Generic Helper Functions
│   │   ├── __init__.py
│   │   └── helpers.py # e.g., calculate_sha256
│   └── main.py        # FastAPI Entry Point
├── .env.example       # Example Environment Variables
├── requirements.txt   # Python Dependencies
└── Dockerfile         # Multi-stage production build

---

💻 Local Development

1. Clone the repository:

   git clone https://github.com/Ambuj123-lab/indian-legal-ai-expert.git
   cd indian-legal-ai-expert

2. Backend Setup:

   cd backend
   python -m venv venv
   source venv/bin/activate  # Or venv\Scripts\activate
   pip install -r requirements.txt

3. Frontend Setup:

   cd frontend
   npm install

4. Environment Variables: Create a .env file in the root using .env.example as a template.

---

⚖️ Disclaimer

This application is an AI-powered educational tool designed to assist with legal research. It does not constitute legal advice. Users should consult a qualified legal professional for critical matters.

---

Developed with ❤️ by Ambuj Kumar Tripathi