Abstract: To overcome the limitations of traditional reconfigurable intelligent surface (RIS), which rely on predefined coding patterns and lack environmental adaptability, this paper proposes a RIS-based health monitoring system, which is empowered by a large language model (LLM) and can be applied for smart home applications. Using the LLM as its central decision-making unit, the system decomposes natural language instructions into subtasks through engineering. The proposed system integrates a YOLOv7-Pose pose estimation network with a Kalman filter algorithm for real-time human thoracic localization. Based on the obtained coordinates, the system dynamically optimizes the RIS codebook to achieve precise electromagnetic wave focusing. Finally, the improved variational mode decomposition algorithm is used to extract respiratory rates from the echo signals and the system controls the smart home devices based on the measured respiratory rate, and sends an alarm message via WeChat. Experimental results demonstrate an average Euclidean positioning error of 0.09m for the human thoracic region, the instruction decomposition success rate of 91%, and a mean absolute error of 2.31 RPM for respiratory rate monitoring. The system establishes a closed-loop framework from perception to response, enabling natural language-driven adaptive electromagnetic control. This approach offers a novel contactless health monitoring solution for smart healthcare and smart homes.