LegalRAG

A Retrieval-Augmented Generation (RAG) system for legal document analysis, built with FastAPI, Pinecone, and Llama 3.3 via Groq Inference.

Overview

LegalRAG is an AI-powered system that retrieves and analyzes legal documents based on user queries. It utilizes vector-based search to provide relevant case law, legal statutes, and precedents in response to legal queries.

Features

Legal Document Retrieval – Uses Pinecone for fast and efficient case law retrieval.
Natural Language Querying – Allows users to search legal documents using plain English.
Session-Based Handling – Supports session-based query tracking for better user experience.
FastAPI Backend – Provides a scalable and efficient REST API.
Concurrency Handling – Manages multiple legal research sessions efficiently.
Llama 3.3 via Groq Inference – High-speed, low-latency language model inference.

---

Installation

1. Clone the repository

git clone https://github.com/arulkumarann/legalRAG.git
cd legalRAG

2. Create a virtual environment

python -m venv venv
# On Windows
venv\Scripts\activate
# On macOS/Linux
source venv/bin/activate

3. Install dependencies

pip install -r requirements.txt

---

Configuration

Create a .env file and add the following variables:

PINECONE_API_KEY=your_pinecone_api_key
INDEX_NAME=your_pinecone_index_name
GROQ_API_KEY=your_groq_api_key

---

Running the Application

Local Development

uvicorn app.main:app --reload

API will be available at http://localhost:8000

Running the Streamlit Interface

streamlit run app.py

The Streamlit interface will be available at http://localhost:8501

---

API Endpoints

1. Root Check

• GET / – Health check endpoint

2. Chat with RAG

• POST /chat – Query legal documents with RAG
  • Body:

    {
      "session_id": "abc123",
      "query": "What are the key points of Smith v. Johnson?"
    }

  • Response:

    {
      "response": "The key points are breach of contract and negligence...",
      "sources": ["case_smith_v_johnson.pdf"]
    }

3. Session Management

• DELETE /session/{session_id} – End a legal research session
• POST /session/reset/{session_id} – Reset a session's state
• GET /sessions/count – Get the total number of active sessions

4. Retrieve Source Documents

• GET /sources/{session_id} – Fetch relevant legal documents for a session

---

Llama 3.3 via Groq Inference

• Inference Speed: <15ms per query response
• Token Throughput: ~500 tokens/sec
• Latency: Low-latency response optimized for real-time queries
• Memory Usage: Efficient memory footprint compared to traditional on-device LLMs
• Scalability: Supports concurrent user queries without degradation

---

Concurrency & Session Handling

This RAG system supports multiple concurrent users by assigning unique session_id values for each session.

• Sessions track user queries to improve context and accuracy.
• Each session has its own vector search scope in Pinecone, ensuring faster retrieval of case-specific documents.
• Automatic cleanup – Sessions can be ended manually (DELETE /session/{session_id}) or reset (POST /session/reset/{session_id}).

---

Deployment

Deploy on Render

1. Build Command:

   pip install -r requirements.txt

2. Start Command:

   gunicorn app.main:app -k uvicorn.workers.UvicornWorker --workers 1 --threads 2 --timeout 120

3. Environment Variables:
   • PINECONE_API_KEY
   • INDEX_NAME
   • GROQ_API_KEY

---

Memory Optimization

• Uses lazy model loading to reduce memory usage.
• Batch processing to prevent memory spikes.
• Garbage collection after query execution.

---

License

This project is open-source under the MIT License.