RAG System - Fire Safety Document Q&A

A Retrieval-Augmented Generation (RAG) system for querying fire safety documents using semantic search, LLM reranking, and structured answer generation.

Features

• Semantic search with FAISS vector database
• Hybrid retrieval (FAISS + BM25)
• LLM-based reranking using Groq
• Structured JSON responses validated with Pydantic
• Automatic extraction of diagrams and images
• REST API with FastAPI
• Docker support

Prerequisites

You need API keys for:

• OpenAI API - For embeddings: https://platform.openai.com/api-keys
• Groq API - For LLM: https://console.groq.com/keys

Quick Start

1. Environment Setup

Create a .env file in the project root:

OPENAI_API_KEY=your_openai_key_here
GROQ_API_KEY=your_groq_key_here

2. Docker Setup

# Build and start
docker-compose up --build

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

Usage

Health Check

curl http://localhost:8000/health

Query API

curl -X POST "http://localhost:8000/chat/answer" \
  -H "Content-Type: application/json" \
  -d '{"question": "What are fire door requirements?"}'

API Documentation

Visit http://localhost:8000/docs for interactive API documentation.

Testing

# Run all tests
pytest

# Run with coverage
pytest --cov=src --cov-report=html

# Lint code
flake8 src/ api.py main.py tests/

# Format code
black src/ api.py main.py tests/

Configuration

Default settings in src/config.py:

• Chunk size: 1000 characters - Balances context preservation with retrieval precision. Large enough to capture complete concepts in technical documentation while small enough for focused retrieval.
• Chunk overlap: 150 characters - Ensures critical information spanning chunk boundaries isn't lost, maintaining continuity across splits.
• Top-k: 5 - Provides sufficient context diversity without overwhelming the LLM. After reranking from 30 FAISS candidates, top 5 ensures quality over quantity.
• Candidate-k: 30 - Initial FAISS retrieval pool for LLM reranking. Large enough to capture diverse relevant passages while keeping reranking latency manageable (~2-3s).
• Embedding: text-embedding-3-small - OpenAI's cost-effective model (1536-dim) with strong semantic understanding. Balances performance and affordability for technical document retrieval.
• Temperature: 0 - Ensures deterministic, factual outputs critical for technical/regulatory documentation. Eliminates hallucination risk in safety-critical domain.

Technologies

• Python 3.12+
• FastAPI - REST API framework
• FAISS - Vector similarity search
• LangChain - RAG components
• OpenAI - Text embeddings
• Groq - LLM (Llama 3.1 8B Instant)
• Loguru - Logging
• Pydantic - Data validation
• Docker - Containerization
• Pytest - Testing