AttorneysInRAGs

A legal compliance analysis tool that automatically checks Terms of Service and Privacy Policy documents against Indian IT laws and regulations using RAG (Retrieval-Augmented Generation).

Overview

AttorneysInRAGs analyzes legal documents through a multi-stage pipeline:

1. Filtering - Extracts relevant legal clauses using ontology-based keyword matching + AI classification
2. Matching - Embeds clauses and queries a vector database of laws to find potential matches
3. Analysis - Uses an LLM to determine if matches constitute actual violations

Architecture

┌─────────────┐    ┌──────────────┐    ┌─────────────┐    ┌──────────────┐
│  Input ToS  │───▶│  Filter      │───▶│  Vector DB  │───▶│  LLM        │
│  Document   │    │  (filter.py)  │    │  (matcher)  │   │  (inference)  │
└─────────────┘    └──────────────┘    └─────────────┘    └──────────────┘
                         │                    │                   │
                   Ontology + AI        ChromaDB +            Ollama +
                   Classification       BGE Embeddings        Mistral

Installation

Prerequisites

• Python 3.11 or 3.12.3 (<3.14)
• Ollama with mistral:latest model

Setup

# Clone and enter directory
cd AttorneysInRAGs

# Create virtual environment
python -m venv venv
source venv/bin/activate

# Install dependencies
pip install chromadb sentence-transformers spacy transformers fastapi uvicorn httpx

# Download spaCy model
python -m spacy download en_core_web_sm

# Pull Ollama model
ollama pull mistral:latest

Initialize Database

python experimentation/db_generator.py

This populates the ChromaDB vector database with embedded law rationales from backend/database/db.json.

Usage

CLI Testing

python backend/main.py

Reads backend/text.txt and runs the full pipeline.

API Server

# From project root
uvicorn backend.api:app --host 0.0.0.0 --port 8000

Endpoint

POST /analyze

curl -X POST http://localhost:8000/analyze \
  -H "Content-Type: application/json" \
  -d '{"text": "Your Terms of Service text here..."}'

Response

{
  "summary": "Executive summary of violations found.",
  "aggregations": {
    "total_violations": 2,
    "critical_severity": 1,
    "high_severity": 1,
    "medium_severity": 0,
    "low_severity": 0
  },
  "violations": [
    {
      "violating_rule": "ToS clause text...",
      "actual_rule": "The actual law text...",
      "source": "[IT_ACT_SEC_43A] DATA_SHARING, LIABILITY",
      "severity": "CRITICAL",
      "reason": "Why this is a violation"
    }
    ...
  ]
}

Project Structure

AttorneysInRAGs/
├── backend/
│   ├── api.py              # FastAPI server
│   ├── main.py             # CLI pipeline runner
│   ├── filter.py         # RelevanceFilter (ontology + AI)
│   ├── matcher.py       # Vector search + matching
│   ├── inference.py         # LLM inference (Ollama)
│   ├── text.txt            # Sample input for testing
│   └── database/
│       ├── db.json         # Law rules database
│       └── chroma_db/      # Vector embeddings
├── experimentation/
│   ├── db_generator.py     # Populate ChromaDB
│   └── svo.py              # Text distillation experiments
└── README.md

Configuration

Embedding Model

Uses BAAI/bge-small-en-v1.5 (384-dim) for fast, high-quality embeddings.

LLM

Default: Ollama with mistral:latest. Configure in backend/inference.py:

OLLAMA_URL = "http://localhost:11434/api/generate"
MODEL = "mistral:latest"

Similarity Threshold

Adjust in backend/matcher.py:

threshold=0.40  # Lower = stricter matching

Law Domains Covered

• DATA_COLLECTION - Data gathering practices
• DATA_RETENTION - Storage duration requirements
• DATA_SHARING - Third-party sharing rules
• CONSENT - User consent requirements
• SECURITY_PRACTICES - Security standards
• BREACH_RESPONSE - Incident notification rules
• USER_RIGHTS - Access, correction, deletion rights
• GRIEVANCE - Complaint handling procedures
• LIABILITY - Liability and indemnification
• SENSITIVE_DATA - Special category data rules
• CHILDREN_DATA - Minor protection rules
• LOGGING_AUDIT - Audit trail requirements

Law documents used

IT Act, 2000 Sections 43 & 66

DPDP Act, 2023

IT Security Practices Rules

CERT-In Guidelines

IT Intermediary Guidelines

License

MIT