A Privacy-Preserving RAG Framework for Diabetes Management Chatbots

This paper presents a privacy-preserving medical chatbot designed to support diabetes education and selfmanagement using open-source artificial intelligence models. The chatbot aims to answer common diabetes-related questions, including questions about blood glucose levels, symptoms, treatment understanding, and lifestyle recommendations. It provides immediate, conversational, and reliable responses to help patients manage this chronic disease. The system integrates an open-source biomedical large-scale language model and a retrieval-enhanced generation (RAG) framework to ensure that generated responses are based on verified medical knowledge from trusted sources such as the World Health Organization (WHO) and the Medical Question Answering Dataset (MedQuAD). The chatbot supports both text and voice interaction and includes a user profile module for context-aware personalized responses. Safety-focused prompts are implemented to prevent medical diagnosis or prescription recommendations and to encourage consultation with healthcare professionals when necessary. The system was evaluated using 100 diabetes-related questions from the MedQuAD dataset using an LLM-as-a-judge evaluation framework. The results achieved an average score of 4.34/5, with 83% of responses rated as medically correct or substantially correct. The findings suggest that combining open-source biomedical LLMs with RAG and privacy-preserving local deployment can provide accessible and reliable educational support for diabetes management.