RAG Improves LLM Safety in Mental Health Chatbots, Study Finds

A study by Wysa, a digital mental health intervention (DMHI) platform, quantifies the incremental safety benefit of Retrieval-Augmented Generation (RAG) in large language models (LLMs). Published on arXiv (2607.24817), the research evaluates six LLM models within Wysa's layered safety architecture, comparing RAG-enabled versus RAG-disabled modes. The system combines rule-based filters, symbolic escalation protocols, and neural classification. Annotated anonymized real and synthetic user-chatbot exchanges, reviewed by a qualified clinician, were used to test intent detection in volatile situations. Results show that RAG significantly enhances accuracy and reduces hallucination, addressing a critical gap in pure parametric LLMs that lack specific safety architecture. The study provides the first controlled quantification of a single safety layer's contribution in a commercial DMHI.