🤖 RAG (Retrieval-Augmented Generation) System

A simple implementation of a Retrieval-Augmented Generation system using OpenAI's GPT-4o, sentence transformers, and FAISS for efficient similarity search.

📖 Overview

This project demonstrates how to build a RAG system that:

• 📝 Splits text into sentences for better granular retrieval
• 🔗 Creates embeddings using SentenceTransformers
• 🗃️ Stores embeddings in a FAISS index for fast similarity search
• 🔍 Retrieves relevant context for user queries
• 🤖 Generates responses using OpenAI's GPT-4o with proper citations

✨ Features

• 🔐 Environment-based configuration - API keys stored securely in .env file
• 📖 Sentence-level retrieval - Fine-grained context retrieval using NLTK tokenization
• ⚡ Efficient similarity search - FAISS indexing for fast vector search
• 📚 Citation support - Generated responses include source citations
• 📊 Confidence scoring - AI-generated confidence scores for answers
• 🚀 Modern OpenAI API - Compatible with OpenAI Python library v1.0+

📋 Prerequisites

• 🐍Python 3.7+
• 🔑OpenAI API key
• 📓Jupyter Notebook or JupyterLab

🚀 Installation

1. 📁Clone or download this repository

2. 📦Install required packages:

   pip install openai sentence-transformers faiss-cpu nltk tiktoken python-dotenv

3. 🔐Set up your environment variables:

   Create a .env file in the project directory:

   OPENAI_API_KEY=sk-your-actual-api-key-here
   

   ⚠️ Important: Replace sk-your-actual-api-key-here with your actual OpenAI API key.

💻 Usage

🔴 Running the Notebook

1. Open the Jupyter notebook: 📓

   jupyter notebook RAG.ipynb

2. ⚡Execute the cells in order:

   • 📦Cell 1: Install dependencies
   • 🔧Cell 2: Import libraries and load environment variables
   • 🔍Cell 3: Process text (split, embed, index)
   • 📝Cell 4: Define retrieval function
   • 🤖Cell 5: Define answer generation function
   • ❓Cell 6: Test with example query

💡 Example Query

query = "What is RAG and how does it reduce hallucinations?"
print(generate_answer(query))

✨Expected Output:

Retrieval-Augmented Generation (RAG) is a method that improves large language models by allowing them to retrieve information from external documents [1]. It reduces hallucinations by first looking up relevant information and then generating a response using that retrieved context [3]. This approach makes the answers more fact-based [2].

Confidence Score: 1

⚙️ How It Works

1. 📝 Text Processing

• Tokenization: Uses NLTK's sentence tokenizer to split documents into sentences 📖
• Embedding: Converts each sentence to a vector using all-MiniLM-L6-v2 model 🔢

2. 🗃️ Indexing

• FAISS Index: Creates an efficient L2 distance-based index for similarity search 🎯
• Storage: Embeddings are stored in memory for fast retrieval 💾

3. 🔍 Retrieval

• Query Embedding: User queries are converted to the same vector space 🔄
• Similarity Search: FAISS finds the most similar sentences 🎯
• Ranking: Results are ranked by cosine similarity distance 📊

4. 🤖 Generation

• Context Assembly: Retrieved sentences are formatted with citation numbers 📝
• Prompt Engineering: Structured prompt guides the AI to cite sources and provide confidence 🎯
• Response Generation: OpenAI GPT-4o generates the final answer ✨

📁 Project Structure

RAG/
├── RAG.ipynb           # Main Jupyter notebook with implementation 📓
├── .env                # Environment variables (API keys) 🔐
├── README.md           # This file 📖
└── requirements.txt    # Python dependencies (optional) 📦

🔧 Key Components

📦 Dependencies

• openai - OpenAI API client for GPT-4o 🤖
• sentence-transformers - For creating sentence embeddings 🔢
• faiss-cpu - Efficient similarity search and clustering 🎯
• nltk - Natural language processing toolkit 📝
• python-dotenv - Environment variable management 🔐
• tiktoken - OpenAI's tokenization library ⚡

🛠️ Main Functions

• retrieve_sentences(query, top_k=4) - Retrieves most relevant sentences 🔍
• format_context(snippets) - Formats retrieved sentences with citations 📝
• generate_answer(query) - Generates final answer with citations and confidence 🤖

🎨 Customization

🔄 Changing the Embedding Model

embedder = SentenceTransformer("your-preferred-model")

⚙️ Adjusting Retrieval Parameters

top_snippets = retrieve_sentences(query, top_k=6)  # Retrieve more context

🎯 Modifying the Prompt

Edit the prompt in the generate_answer function to change response style or requirements.

🔒 Security Notes

• 🔐 API keys are stored in .env file (not in code)
• 📝 Add .env to your .gitignore file
• ⛔ Never commit API keys to version control
• 🏭 Use environment variables in production

🐛 Troubleshooting

⚠️ Common Issues

1. NLTK Data Missing 📚

   LookupError: Resource punkt_tab not found
   

   Solution: The notebook automatically downloads required NLTK data ✅

2. OpenAI API Error 🤖

   APIRemovedInV1: You tried to access openai.ChatCompletion
   

   Solution: Code uses the new OpenAI v1.0+ API syntax ✅

3. Missing API Key 🔑

   ValueError: OPENAI_API_KEY not found in environment variables
   

   Solution: Check your .env file and ensure the API key is correct ✅

🚀 Performance Tips

• For larger documents, consider chunking strategies 📄
• Use GPU-accelerated FAISS for better performance with large datasets ⚡
• Consider caching embeddings for frequently accessed documents 💾

📄 License

This project is open source and available under the MIT License.

🤝 Contributing

Feel free to submit issues, feature requests, or pull requests to improve this RAG implementation.

🙏 Acknowledgments

• OpenAI for GPT-4o 🤖
• Sentence Transformers for embedding models 🔢
• FAISS for efficient similarity search 🎯
• NLTK for text processing 📝