VedicGPT - RAG-Based Question Answering System

A Retrieval-Augmented Generation (RAG) system that uses Cohere LLM to answer questions about Vedic scriptures with source citations.

Features

• 📚 Loads all text/markdown files from the data folder
• 🔍 Uses FAISS for efficient vector search (cached on disk)
• 🧠 Supports Cohere or local SentenceTransformer embeddings
• 🤖 Powered by Cohere's Command-R-Plus model for answers
• 💬 Interactive chat interface with Gradio
• 📎 Automatic source citation for all answers
• 🕉️ Optimized for Vedic scripture queries

Setup

1. Install Dependencies

pip install -r requirements.txt

2. Get Cohere API Key

1. Visit Cohere Dashboard
2. Sign up or log in
3. Navigate to API Keys section
4. Copy your API key

3. Configure Environment

Edit the .env file and add your Cohere API key (needed for chat responses):

COHERE_API_KEY=your-actual-api-key-here

Optional environment variables:

# Use Cohere for embeddings (requires higher rate limits) or keep default local model
EMBEDDING_PROVIDER=local   # options: local, cohere
LOCAL_EMBED_MODEL=sentence-transformers/all-MiniLM-L6-v2
COHERE_CHAT_MODEL=command-r-plus-08-2024
COHERE_EMBED_MODEL=embed-english-v3.0

4. Add Your Data

Place your Vedic text files (.txt or .md) in the data folder. The system will automatically:

• Load all documents
• Split them into chunks
• Create a vector database
• Enable semantic search

Usage

Run the Application

python app.py

The application will:

1. Load all documents from the data folder
2. Create a vector store for semantic search
3. Launch a web interface at http://localhost:7860

Chat Interface

1. Open your browser to http://localhost:7860
2. Type your question in the text box
3. Click "Send" or press Enter
4. Receive answers with source citations

Example Questions

• "What is the meaning of dharma in the Bhagavad Gita?"
• "Explain the concept of karma"
• "What does the Gita say about meditation?"

How It Works

1. Document Loading: All .txt and .md files from the data folder are loaded
2. Text Chunking: Documents are split into manageable chunks with overlap
3. Embedding: Each chunk is converted to vector embeddings (Cohere or local models)
4. Vector Store: FAISS indexes the embeddings for fast retrieval and is cached on disk
5. Query Processing: User questions are embedded and matched against the database
6. Context Building: Relevant chunks are retrieved and formatted
7. LLM Response: Cohere generates an answer based on the context
8. Source Citation: The response includes references to source documents

Project Structure

VedicGPT/
├── app.py                 # Main RAG application
├── ocr.py                 # PDF to text conversion
├── data/                  # Your text documents
├── requirements.txt       # Python dependencies
├── .env                   # Environment variables (API keys)
└── README.md             # This file

Technologies Used

• Cohere: LLM for question answering
• LangChain: Framework for RAG pipeline
• FAISS: Vector database for semantic search
• Gradio: Web interface for chat
• PyMuPDF: PDF processing
• Tesseract: OCR for scanned documents

Tips

• For better results, ensure your documents are well-formatted
• Use the OCR script (ocr.py) to convert PDF files to text
• The system works best with clear, structured text
• Adjust chunk_size and k (number of retrieved docs) for optimization

Troubleshooting

Error: "No module named 'cohere'"

pip install -r requirements.txt

Error: "Invalid API key"

• Check your .env file has the correct Cohere API key
• Ensure there are no extra spaces or quotes

Error: "TooManyRequestsError" when building embeddings

• Leave EMBEDDING_PROVIDER=local to avoid Cohere embedding rate limits
• Upgrade to a production Cohere key if you must use Cohere embeddings
• Delete the storage/vectorstore folder if you need to rebuild the cache

No documents loaded

• Verify files exist in the data folder
• Check file extensions are .txt or .md
• Ensure files have UTF-8 encoding

License

MIT License