Doc-RAG: Document Retrieval-Augmented Generation System

Overview

Doc-RAG is a specialized question-answering system that uses Retrieval-Augmented Generation (RAG) to provide accurate answers about technical software documentation. The system currently supports two tools:

• ArchiCAD
• RFEM

Architecture

The system consists of three main components:

1. Frontend Service (Port 9000)

   • Web interface for user interactions
   • Built with Flask and simple HTML/CSS
   • Communicates with backend service via REST API

2. Backend Service (Port 8000)

   • Core RAG implementation
   • Handles document embedding and retrieval
   • Integrates with OpenAI for response generation
   • Uses Pinecone for vector storage

3. ETL Pipeline

   • Processes PDF documentation
   • Generates embeddings using sentence-transformers
   • Indexes documents in Pinecone database

Prerequisites

• Python 3.9
• Docker and Docker Compose
• Pinecone API credentials
• OpenAI API key

Running the application

1. Create a .env file in the root directory with:

PINECONE_API_KEY=your_pinecone_key
PINECONE_ENV=your_pinecone_environment
OPENAI_API_KEY=your_openai_key

2. Run using docker compose:

docker compose up --build

This will start the frontend and backend services. The frontend can then be accessed at http://localhost:9000.

Adding new Documents

To add new documentation:

1. Place PDF documentation in the docs directory
2. Run the ETL process:

python3 etl.py <pdf_path> <tool_name>

Where tool_name is currently either archicad or rfem. This will extract the text from the PDF and embed it using sentence-transformers. The embeddings are then stored in Pinecone.

System Components

Frontend

• Simple web interface with tool selection
• Chat-like interface for questions and answers

Backend

• API endpoint for processing queries
• Vector similarity search using Pinecone
• Context-aware response generation using GPT-3.5-turbo
• Sentence transformer embeddings (all-MiniLM-L6-v2)

ETL Pipeline

• PDF text extraction using PyMuPDF
• Text chunking with sentence splitting
• Document embedding generation
• Vector database indexing