부산대학교 도서관

This paper discusses the challenges of using Large Language Models (LLMs) in medical chatbots for chronic disease self-management. Accordingly, we define an architecture specifically devised to deal with issues related to reliability, clinical trials, and privacy. Two solutions are compared to prevent data disclosure: a filtering mechanism for sensitive data with an external LLM, and a locally deployed LLM using open-source models. Experimental results underscore the challenges in effectively instructing the local LLM so as to provide performances comparable to GPT-3.5.