Retrieval-augmented Generation of Enhanced Trigger-action Programming Rules in Smart Home

Trigger-action programming (TAP) is a popular paradigm for smart home automation, enabling users to create rules in form of “IF trigger, THEN action”. While large language models (LLMs) offer a promising path for generating TAP rules from natural language, their vanilla application, such as relying solely on pre-trained knowledge and basic prompting, falters as platforms evolve to support enhanced TAP rules. Such rules incorporate scripting for conditional logic, computations, and external API calls. Enhanced TAP rules demand users to express complex logic and environmental context, making the creation of such rules difficult without a strong programming background. This paper introduces HomeGenii, a retrieval-augmented generation (RAG) system that automates enhanced TAP creation. HomeGenii constructs a compact yet representative rulebase, retrieves semantically aligned rules using a cluster-then-search approach, and applies compression techniques to minimize token overhead. Evaluation shows HomeGenii improves enhanced TAP rule generation accuracy to 84%, a 70% increase over systems without RAG. Our work demonstrates a viable pathway for enabling non-expert users to leverage LLMs for expressive and complex home automation.