BioMistral Medical RAG Chatbot

Project Overview

• Goal: Develop a chatbot to answer user queries about health concerns, symptoms, treatments, and more.
• Purpose: Serve as a reliable and accessible resource for medical information and advice.
• Learning Context: This project was completed as part of the learning stages of the GUVI SAWIT.AI Women-Only, Gen AI Learning Challenge.

Key Components

1. Llama Library: Utilized for working with language models.
2. Retriever: Gathers context-relevant information to enhance LLM responses.
3. Prompt Template: Combines retriever results and user queries for accurate response generation.

Technologies Used

• Langchain: For building the chatbot framework.
• Hugging Face Transformers: For implementing the LLM and embedding model.
• PyPDF2: For loading and parsing PDF documents.

Implementation

• Retrieval Augmented Generation (RAG) Chain:
  • Integrates the retriever, LLM, and prompt template.
  • Facilitates dynamic retrieval and generation of medical information.

Development Steps

1. Environment Setup

• Set up the development environment on Google Colab.
• Install required Python packages (Langchain, Sentence Transformers, Hugging Face, etc.).

2. Document Preprocessing

• Import medical documents into the project environment.
• Extract text using libraries like PyPDF2 for PDFs and docx for Word files.
• Utilize Langchain’s text splitter to chunk the text into manageable segments for retrieval.

3. Creating Embeddings and Vector Store

• Generate embeddings for text chunks using a pre-trained embedding model from Hugging Face.
• Create a Chroma vector store for efficient storage and retrieval of embeddings.
• Index text chunks with their corresponding embeddings for similarity search.

4. LLM Integration

• Load a pre-trained LLM using the Llama library for generating informative responses.
• Design a prompt template combining retrieved context and user queries.
• Construct a Retrieval Augmented Generation (RAG) chain using Langchain’s Chain Class to integrate the retriever, LLM, and prompt template.

Testing

• Conducted testing with various medical-related prompts.

Models and Data used

• https://huggingface.co/NeuML/pubmedbert-base-embeddings
• https://huggingface.co/MaziyarPanahi/BioMistral-7B-GGUF/tree/main
• https://www.nhlbi.nih.gov/files/docs/public/heart/healthyheart.pdf