RAGChatbot

Experimental project with OpenAI LLM + RAG using Huggingface + Pinecone vectorstore

PDF AgentForce is an AI-powered chatbot designed to extract, index, and query information from uploaded PDF documents. It integrates advanced natural language processing techniques to provide document-centric responses, as well as fallback capabilities to open-ended GPT-based search when document context is insufficient.

Key Features

1. Pinecone Integration for Vector Search: Uses Pinecone to store and retrieve document embeddings, ensuring fast and efficient similarity search.
2. OpenAI GPT-4 Integration: Combines document-based answers with GPT-4 to provide general knowledge or fallback responses when documents don’t contain the relevant information.
3. Hugging Face Integration: Downloads and processes documents from a Hugging Face repository upon initialization.
4. Streamlit Interface: An intuitive UI that allows users to upload PDF documents, query them, and view responses interactively.

Technology Stack

• Pinecone: Used for storing and querying embeddings with high-performance vector search.
• OpenAI GPT-4: Powers natural language responses and fallback search capabilities.
• Streamlit: A lightweight and powerful framework used to build the app's user interface.
• Hugging Face Hub: Used for storing, downloading, and embedding PDF documents for document-based answers.
• Sentence Transformers: Utilized for generating embeddings for both the document content and user queries.

How It Works

1. Document Upload: Users can upload PDFs which are processed to extract text using the pdf_utils module.
2. Text Embedding: The extracted text is converted into embeddings using the Sentence Transformers model and stored in Pinecone.
3. Querying: When a query is made, the bot searches through the stored document embeddings in Pinecone to retrieve the most relevant information.
4. GPT Fallback: If the relevant context is not found in the document, the bot falls back to OpenAI GPT-4 to answer the query.
5. Response Streaming: Responses are streamed in real-time, offering a seamless user experience.

Project Structure . ├── app_chatbot.py # The main application file for Streamlit ├── utils/ │ ├── pdf_utils.py # Functions for PDF text extraction │ ├── huggingface_utils.py # Utility for uploading to Hugging Face │ ├── embedding_utils.py # Embedding generation using Sentence Transformers │ ├── pinecone_utils.py # Pinecone initialization and vector store setup │ ├── huggingface_download_embed_utils.py # Functions for downloading from Hugging Face └── README.md # Project documentation

install required libraries

pip install -r requirements.txt

Add the necessary environment variables:

PINECONE_API_KEY: Your Pinecone API key. PINECONE_INDEX_NAME: The name of the Pinecone index. OPENAI_API_KEY: Your OpenAI API key. HUGGINGFACE_API_KEY: Your Hugging Face API key.

Run the application: streamlit run app_chatbot.py

Future Enhancements Advanced Document Parsing: Implement more robust parsing for a wider variety of document formats - Multimodal. Improved Query Accuracy: Enhance the embedding and search mechanisms to deliver more precise answers.

Contributions This is purely an experimental and hobby work! But I welcome contributions to improve this project!