🤖 RAG System with Cohere

This repository contains implementations of two Retrieval-Augmented Generation (RAG) systems using the Cohere API. The project provides both a simple RAG implementation and an enhanced contextual RAG implementation to demonstrate different approaches and performance characteristics.

📚 Overview

The project implements two RAG architectures:

1. Simple RAG 🔍: An implementation following the basic RAG pattern - document loading, chunking, embedding, retrieval, and generation.

2. Contextual RAG 🧠: A more sophisticated implementation with additional features:

   • Context generation for document chunks
   • Query reformulation based on conversation history
   • Enhanced document retrieval with contextual awareness

Both systems leverage Cohere's API for embeddings and language models, and process PDF documents as knowledge sources.

🛠️ Setup

1. Clone the repository:

   git clone https://github.com/your-username/rag-system-cohere.git
   cd rag-system-cohere

2. Create a virtual environment and install dependencies:

   python -m venv venv
   source venv/bin/activate  # On Windows: venv\Scripts\activate
   pip install -r requirements.txt

3. Create a .env file in the root directory with your Cohere API key:

   COHERE_API_KEY=your_cohere_api_key_here
   

API keys can be obtained by signing up at cohere.ai.

🖥️ Usage

The system can be used in two modes: single query mode or interactive mode.

Command Line Interface

Simple RAG 🔍

# Single query mode
python app.py simple --pdf_path path/to/your/document.pdf --query "Your question here"

# Interactive mode
python app.py simple --pdf_path path/to/your/document.pdf --interactive

# Process a directory of PDFs
python app.py simple --pdf_path path/to/pdf/directory --interactive

Contextual RAG 🧠

# Single query mode
python app.py contextual --pdf_path path/to/your/document.pdf --query "Your question here"

# Interactive mode
python app.py contextual --pdf_path path/to/your/document.pdf --interactive

# Process a directory of PDFs
python app.py contextual --pdf_path path/to/pdf/directory --interactive

Both implementations can be tested to determine which best fits specific use cases.

📊 Comparison Notebook

The included Jupyter notebook comparison.ipynb provides a detailed comparison between the Simple and Contextual RAG implementations. The notebook:

• Loads identical document sets into both systems
• Executes benchmark queries through both systems
• Compares response quality and relevance
• Analyzes performance differences between approaches
• Visualizes results for comparative analysis

To run the notebook:

jupyter notebook comparison.ipynb

✨ Key Features

Simple RAG 🔍

• Document processing pipeline with PyPDF for loading PDFs
• Text chunking with RecursiveCharacterTextSplitter
• Vector storage with Chroma
• Basic retrieval using similarity search
• Answer generation using retrieved context

Contextual RAG 🧠

• Enhanced document processing with context generation
• Context summaries created for each document chunk
• Conversation-aware query reformulation
• Improved retrieval with contextual information
• Rate limiting to manage API usage

🔧 Technical Details

Document Processing 📄

Both systems process PDF documents by:

1. Loading the PDF using PyPDFLoader
2. Splitting the text into manageable chunks
3. Storing chunks in a vector database (Chroma)

The Contextual RAG system adds an additional step of generating contextual summaries for each chunk using Cohere's API.

Query Processing 🔎

• Simple RAG: Directly uses the user's query for retrieval
• Contextual RAG: Reformulates queries based on conversation history for improved context awareness

Configuration ⚙️

The config.py file contains configurable parameters:

• API keys
• Vector database settings
• Text chunking parameters
• Default retrieval settings

Performance Considerations ⏱️

• The Contextual RAG system makes more API calls and has higher latency due to the additional context generation step.
• The Simple RAG system is faster but may lack contextual awareness in multi-turn conversations.
• Rate limiting has been implemented in the Contextual RAG system to avoid API throttling.

⚖️ System Comparison

The two RAG systems offer different trade-offs:

• Simple RAG 🔍: Provides a lightweight, faster implementation suitable for straightforward question-answering tasks.
• Contextual RAG 🧠: Delivers more contextually relevant answers, especially in multi-turn conversations, at the cost of increased latency and API usage.

Selection between implementations depends on specific use case requirements regarding response quality, conversation handling, and performance constraints.

🚀 Future Development

Potential enhancements:

• 📚 Integration with additional document types (beyond PDFs)
• 🔄 Implementation of hybrid retrieval techniques
• ⚡ Caching mechanisms to reduce API calls
• 🔧 Fine-tuning of chunking and retrieval parameters
• 🖥️ UI implementation for improved user interaction

Contributions welcome for any of these enhancements or other improvements.