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

Overview

Trigger-action programming (TAP) provides a declarative framework for smart home automation rule specification through conditional IF-THEN statements. Contemporary TAP platforms support enhanced rule variants incorporating embedded scripting, conditional logic, computational operations, and external API integrations. However, large language models applied through standard prompting techniques demonstrate insufficient performance for generating such complex rules, as they rely on pre-trained knowledge without platform-specific context. This work presents HomeGenii, a retrieval-augmented generation system designed to enhance LLM-based rule generation by incorporating domain-specific rule repositories and semantic similarity matching.

Methods and approach

HomeGenii implements a three-stage architecture: (1) construction of a compact yet representative rulebase from the target smart home platform, (2) retrieval of semantically aligned exemplar rules using a cluster-then-search strategy to optimize retrieval efficiency, and (3) application of token compression techniques to reduce context window overhead when augmenting LLM prompts. The system leverages semantic similarity matching to identify contextually relevant rules that serve as in-context examples for generation tasks. The cluster-then-search approach segments the rulebase into semantic clusters prior to executing similarity-based retrieval, thereby reducing computational cost while maintaining retrieval quality.

Key Findings

Empirical evaluation demonstrates that HomeGiii achieves 84% accuracy in generating enhanced TAP rules from natural language specifications. This represents a 70 percentage point improvement relative to LLM systems without retrieval augmentation, establishing the efficacy of domain-specific exemplar grounding. The compression techniques successfully minimize token overhead associated with augmented context while maintaining generation quality, supporting deployment in token-constrained scenarios.

Implications

The findings establish retrieval-augmentation as a viable mechanism for improving LLM performance on domain-specific automation tasks requiring complex logical expressions and platform-specific syntax. By grounding generation in exemplar rules from the target platform, the system circumvents limitations of pre-trained knowledge and enables more accurate synthesis of non-trivial conditional logic. The approach maintains accessibility for non-specialist users while supporting expressive rule specifications previously requiring programming expertise. The cluster-then-search retrieval strategy presents a generalizable pattern for efficient semantic matching in domain-specific corpora, with potential applicability across automation, configuration, and specification generation tasks.

Scope and limitations

This summary is based on the study abstract and available metadata. It does not include a full analysis of the complete paper, supplementary materials, or underlying datasets unless explicitly stated. Findings should be interpreted in the context of the original publication.