Ashraful Islam Paran

Ashraful Islam Paran

ChatWithPDF-Rag-App

Font Awesome Icons

View on GitHub | Live Demo

Overview

ChatWithPDF-Rag-App is an intelligent document interaction system that allows users to have natural conversations with PDF documents. The application uses RAG (Retrieval-Augmented Generation) technology to provide accurate, context-aware responses to questions about PDF content.

Application Interface

Key Features

  • PDF Question Answering: Ask questions about any uploaded PDF document
  • Custom Prompt Engineering: Uses carefully crafted prompts for accurate and context-aware responses
  • Source Document Tracking: Displays the source documents used to generate each answer
  • Advanced Language Models:
    • Uses sentence-transformers/all-MiniLM-L6-v2 for embeddings
    • Powered by mistralai/Mistral-7B-Instruct-v0.3 for language generation

Technical Details

  • Backend: Python 3.11+
  • Frontend: Streamlit
  • Vector Database: FAISS for efficient similarity search
  • Authentication: Hugging Face API integration
  • Duration: Mar 2024 – Present

Installation and Setup

  1. Clone the repository:
    git clone https://github.com/ashrafulparan2/ChatWithPDF-Rag-App.git
cd ChatWithPDF-Rag-App
  2. Install dependencies:
    pip install -r requirements.txt
  3. Set up Hugging Face authentication:
    huggingface-cli login
  4. Run the application:
    streamlit run app.py --server.enableCORS false --server.enableXsrfProtection false

Implementation Details

Vector Store Setup

  • The application uses FAISS (Facebook AI Similarity Search) for efficient document retrieval
  • Requires pre-generated vectorstore files (index.faiss and index.pkl) in the vectorstore/db_faiss/ directory
  • Embeddings are generated using the sentence-transformers/all-MiniLM-L6-v2 model

Question Answering Pipeline

  1. User uploads a PDF document
  2. Document is processed and vectorized
  3. User questions are embedded and matched against the document vectors
  4. Relevant context is retrieved and fed to the Mistral-7B model
  5. Model generates natural, context-aware responses

User Interface

  • Clean, intuitive Streamlit interface
  • Drag-and-drop PDF upload
  • Real-time question answering
  • Source document display for transparency
  • File size limit of 200MB per PDF