The use of large language models in mental health chatbots

Abstract

This project explores the integration of Large Language Models (LLMs) into mental health chatbots to provide accessible, empathetic, and supportive interactions for individuals in Singapore facing psychological distress. Given the rising prevalence of mental health issues and societal stigma surrounding help-seeking behavior, LLM-powered chatbots offer a discreet platform for emotional support. The study investigates the effectiveness of Retrieval-Augmented Generation (RAG) frameworks in improving the contextual relevance, groundedness, and accuracy of chatbot responses. Multiple retrieval methods—including FAISS, BM25, and a hybrid approach—were evaluated using a custom-built dataset and prompt engineering techniques tailored for therapeutic dialogue. The hybrid method demonstrated superior performance in context relevancy, highlighting the benefits of combining lexical and semantic search. While results show promise in generating emotionally intelligent and well-grounded responses, limitations such as narrow dataset scope and vector indexing inefficiencies were noted. The report concludes with recommendations for future enhancements through supervised fine-tuning and conversational memory, aiming toward the development of more robust, safe, and compassionate AI mental health tools.