PDF RAG Chatbot

Overview

PDF RAG Chatbot is a Streamlit-based AI assistant that allows users to upload PDF documents and interactively ask questions about their content. The chatbot leverages Retrieval-Augmented Generation (RAG) techniques, combining semantic chunking, vector search, and a large language model (LLM) to provide accurate answers based on the uploaded PDF.

Features

• PDF Upload & Processing: Users can upload PDF files, which are semantically chunked and indexed for efficient retrieval.
• Vietnamese Language Support: Uses Vietnamese-specific embedding and LLM models for best results on Vietnamese documents.
• Chat Interface: Intuitive chat UI for asking questions and receiving answers based on the PDF content.
• Session Management: Maintains chat history and allows users to clear conversations.
• Model Loading: Loads embeddings and LLM only once per session for efficiency.

How It Works

1. Upload PDF: The user uploads a PDF file via the sidebar.
2. Processing: The PDF is split into semantic chunks using SemanticChunker and embedded using HuggingFaceEmbeddings.
3. Vector Store: Chunks are stored in a Chroma vector database for fast retrieval.
4. RAG Pipeline: When a question is asked, relevant chunks are retrieved and passed to the LLM (Vicuna-7B) using a RAG prompt.
5. Chat: The assistant answers questions based on the PDF content, maintaining a chat history.

Technologies Used

• Streamlit: For the web UI and session management.
• LangChain: For document loading, chunking, and RAG pipeline.
• HuggingFace Transformers: For embeddings and LLM (Vicuna-7B).
• Chroma: For vector database and retrieval.
• PyPDFLoader: For PDF parsing.

File Structure

• app.py: Main Streamlit app, UI, and chat logic.
• model.py: Loads embeddings and LLM models.
• pdf.py: Handles PDF processing, chunking, and RAG pipeline setup.
• session_state.py: Initializes and manages Streamlit session state variables.
• data/: Contains sample PDF files.
• logo/: Contains logo image for the app.

Setup Instructions

1. Install Dependencies:
   • Python 3.11+
   • Install required packages:

     pip install streamlit langchain langchain_huggingface langchain_chroma langchain_community langchain_experimental transformers torch chromadb
     

2. Run the App:

   streamlit run app.py
   

3. Usage:
   • Upload a PDF file in the sidebar.
   • Click "Xử lý PDF" to process the document.
   • Ask questions in the chat box.
   • Clear chat history or upload a new PDF as needed.

Customization

• Models:
  • Embeddings: bkai-foundation-models/vietnamese-bi-encoder
  • LLM: lmsys/vicuna-7b-v1.5
• Chunking:
  • Uses semantic chunking for better context retrieval.

Notes

• The app is optimized for Vietnamese documents but can be adapted for other languages by changing the models.
• Large models may require significant system resources (RAM/VRAM).

License

This project is for educational and research purposes.

Credits

• BKAI Foundation Models
• Vicuna LLM
• LangChain
• Streamlit