A Trustworthy RAG Approach for Healthcare and Health Tourism

Abstract

While health tourism requires accurate and verifiable information for international patients, generative models often suffer from hallucinations and weak source grounding. To address these limitations, we develop a specialized Retrieval-Augmented Generation (RAG) chatbot for healthcare support and health tourism assistance. The system is built on a curated knowledge base of 268,119 validated question-answer pairs spanning 12 specialized domains, and employs hybrid dense-sparse retrieval with BGE-M3 on Qdrant, orchestrated by LangGraph and powered by Qwen3-14B for response generation. In a preliminary evaluation with 80 test queries, the system achieved 93.5% overall intent accuracy, with an average response time of 65.8 s and 43.2 s time-to-first-token (TTFT), indicating reliable intent routing and responsive streaming behavior across heterogeneous user scenarios. © 2026 IEEE.