Safety-Constrained LLM System Designed for Public Health Information

This paper details the design and implementation of a safety-constrained large language model (LLM) system tailored for public health information access, with a specific focus on maternal and child health (MCH) resource navigation. While LLMs offer intuitive interfaces for information retrieval, their deployment in critical sectors like healthcare carries inherent risks related to safety, trust, and the potential for uncontrolled or inaccurate generation. This work explores practical architectural patterns to mitigate these risks. The proposed system employs a multi-layered architecture. Key components include domain-restricted Retrieval-Augmented Generation (RAG), which grounds all responses in curated public health resources, thereby avoiding reliance on the LLM's general pre-trained medical knowledge. Strict boundary enforcement mechanisms are in place to prevent the system from offering medical advice. Additionally, anonymous multi-user session management and comprehensive audit logging are integrated for monitoring and compliance purposes. Scenario-based validation, covering in-scope, out-of-scope, and emergency queries, demonstrated consistent enforcement of safety constraints, reliable resource grounding, and stable system performance with an average response time of 5.3 seconds. The findings offer practical guidance for deploying LLM-based systems in healthcare and other domains where stringent information boundaries and accountability are paramount.