LEGAL AI - Legal Document Retrieval & Virtual Legal Assistant

AI-powered legal assistant for Indian law using RAG (Retrieval-Augmented Generation). Search across IPC, BNS 2023, Constitution, CrPC, 24,000+ legal Q&A pairs, and 146,000+ court judgments.

Application Results

Legal Search — High Confidence Result

Personal Legal Assistant

Architecture

┌─────────────────────────────────┐
│   Next.js Frontend (3D UI)      │  ← localhost:3000
│   Three.js + Tailwind + SSE     │
└──────────────┬──────────────────┘
               │ HTTP / SSE
               ▼
┌─────────────────────────────────┐
│   FastAPI Backend                │  ← localhost:8000
│                                  │
│   ┌───────────────────────────┐ │
│   │  Retriever (pgvector)     │ │
│   │  Cross-encoder reranker   │ │
│   │  Gemini 2.5 Flash (LLM)  │ │
│   └───────────────────────────┘ │
└──────────────┬──────────────────┘
               │ PostgreSQL
               ▼
┌─────────────────────────────────┐
│   Neon (pgvector)               │  ← Cloud database
│   384-dim embeddings            │
│   164K+ legal document chunks   │
└─────────────────────────────────┘

AI/ML Technologies Used

Category	Technology	Purpose
LLM	Google Gemini 2.5 Flash	Answer generation, query expansion
Embeddings	sentence-transformers (all-MiniLM-L6-v2)	384-dim text embeddings (BERT-based)
Reranker	cross-encoder/ms-marco-MiniLM-L-6-v2	Cross-encoder learning-to-rank model
Deep Learning	PyTorch, Transformers	Model inference for embeddings and reranking
Vector Database	Neon PostgreSQL + pgvector	Semantic similarity search over 164K vectors
NLP	tiktoken	Token counting and text chunking
PDF Parsing	PyMuPDF (fitz)	Extract text from legal PDF documents
Data Processing	pandas, NumPy	Dataset loading and processing
Frontend	Next.js 15, React 19	Web application framework
3D Graphics	Three.js, React Three Fiber, Drei	Interactive 3D animated background
Styling	Tailwind CSS, Framer Motion	UI styling and animations
Backend	FastAPI, Uvicorn	REST API + SSE streaming server
Streaming	Server-Sent Events (SSE)	Real-time token-by-token response streaming

Techniques Used

Technique	Implementation	Details
RAG (Retrieval-Augmented Generation)	Full pipeline	Query → Embed → Retrieve → Rerank → Generate
Semantic Search	pgvector cosine similarity	384-dim vector search over 164K chunks
Cross-Encoder Reranking	ms-marco-MiniLM-L-6-v2	Learning-to-rank for precision (21ms/pair)
Text Classification	Keyword-based classifier	93.3% accuracy, 0.957 Macro F1 across 6 categories
Query Expansion	Gemini-powered	Generate alternative phrasings for better recall
Conversation Memory	Context windowing	Last 5 Q&A turns passed to LLM
Recommendation	Related questions	Suggest follow-up queries from retrieved context
Text Chunking	tiktoken-based	500-token chunks with 100-token overlap

Evaluation Metrics

Category Classifier Performance

Metric	Score
Accuracy	93.3%
Macro Precision	0.972
Macro Recall	0.952
Macro F1	0.957
Categories	6 (Criminal, Civil, Property, Consumer, Labor, Family)
Test Samples	30

Per-Class Precision / Recall / F1

Retrieval Metrics

Metric	Score
Precision@1	13.3%
Precision@5	36.7%
Recall@10	36.7%
MRR	0.216
Avg Search Latency	250ms (164K vectors, no index)

Note: Retrieval precision measures exact source matching. The system often returns correct legal information from alternative sources (e.g., court judgments citing IPC sections instead of IPC.pdf directly), so actual answer quality is significantly higher than source-level precision suggests.

Pipeline Performance

Component	Throughput / Latency
Embedding	59 chunks/sec
Reranking	48 pairs/sec (21ms/pair)
Vector Search	~250ms over 164K vectors
Streaming	Real-time token-by-token via SSE

Dataset Distribution

Latency Breakdown

Features

• Legal Search: Ask any question about Indian law with cited sources and confidence scoring
• Personal Legal Assistant: Describe a problem, get rights + applicable laws + recommended actions
• Streaming Responses: Real-time token-by-token answer generation via SSE
• 3D Interactive UI: Animated star field and floating orbs (Three.js)
• Conversation Memory: Multi-turn chat with 5-turn context window
• Query Expansion: Gemini generates alternative phrasings for better retrieval
• Related Questions: Clickable follow-up suggestions after each answer
• Dark/Light Theme: Toggle with localStorage persistence
• Mobile Responsive: Slide-out drawer sidebar on mobile
• Copy & Export: Copy answers to clipboard or export as PDF
• Keyboard Shortcuts: ⌘K focus, ⌘1/⌘2 switch modes, Escape clear

Data Sources

Document	Chunks	Type
Supreme Court & High Court Judgments	146,459	Case Law
Indian Law QA Dataset	13,296	Q&A Pairs
Companies Act 2013	903	Statute PDF
CPC 1908	850	Statute PDF
Constitution of India	675	Statute PDF
Motor Vehicles Act 1988	363	Statute PDF
IPC, BNS, NDPS, IT Act, Contract Act, etc.	1,544	Statute PDFs
Total	164,090	33 unique sources

Setup

1. Backend

cd legal-rag
pip install -r requirements.txt
cp .env.example .env
# Edit .env with your GEMINI_API_KEY and DATABASE_URL

2. Database (Neon)

Create a free project at neon.tech, then run:

python3 -c "
import psycopg2, os
from dotenv import load_dotenv
load_dotenv()
conn = psycopg2.connect(os.getenv('DATABASE_URL'))
cur = conn.cursor()
cur.execute('CREATE EXTENSION IF NOT EXISTS vector')
cur.execute('''CREATE TABLE IF NOT EXISTS legal_chunks (
  id bigserial PRIMARY KEY, text text NOT NULL,
  embedding vector(384), source text NOT NULL,
  collection text NOT NULL, chunk_hash text UNIQUE NOT NULL,
  metadata jsonb DEFAULT \\'{}\\', created_at timestamptz DEFAULT now())''')
cur.execute('CREATE INDEX IF NOT EXISTS legal_chunks_collection_idx ON legal_chunks (collection)')
conn.commit()
print('Done')
conn.close()
"

3. Ingest Data

python3 ingest/embedder.py

4. Run

bash start.sh

Or use VS Code / Cursor Run and Debug → "Full Stack (Backend + Frontend)"

Open http://localhost:3000

5. Evaluate

python3 evaluate.py
python3 generate_dashboard.py

Project Structure

├── legal-rag/
│   ├── api/main.py              # FastAPI backend (REST + SSE streaming + sessions)
│   ├── ingest/
│   │   ├── csv_loader.py         # Load + chunk CSV Q&A dataset
│   │   ├── pdf_loader.py         # Extract + chunk PDFs (PyMuPDF)
│   │   ├── jsonl_loader.py       # Load + chunk JSONL datasets
│   │   └── embedder.py           # Embed chunks + store in pgvector
│   ├── rag/
│   │   ├── retriever.py          # Semantic search (pgvector) + reranker
│   │   ├── generator.py          # Gemini answer generation with citations
│   │   └── pipeline.py           # Full RAG pipeline
│   ├── assistant/
│   │   └── legal_assistant.py    # Personal legal assistant mode
│   ├── config.py
│   └── requirements.txt
├── frontend/
│   ├── src/app/page.tsx          # Main UI (search + assistant modes)
│   ├── src/components/
│   │   ├── Scene3D.tsx           # Three.js 3D background
│   │   ├── ConfidenceBadge.tsx
│   │   ├── SourceChips.tsx
│   │   └── TypingIndicator.tsx
│   └── src/lib/api.ts            # API client + SSE streaming
├── charts/                       # Evaluation dashboards and metrics
├── evaluate.py                   # Evaluation pipeline
├── generate_dashboard.py         # Performance dashboard generator
├── start.sh                      # Start both servers
└── README.md

Disclaimer

This is AI-generated legal information, not a substitute for professional legal advice. Always consult a qualified lawyer for your specific legal situation.

License

MIT License — see LICENSE for details.