Trustworthy Natural-Language Analytics Over Private Smart Home IoT Data via Time-Aware RAG

Natural-language question answering over private Internet of Things (IoT) data is a promising interface for smart home analytics, but existing large language model (LLM) assistants remain unreliable because they cannot directly access sensor logs, and conventional retrieval-augmented generation (RAG) does not adequately embed temporal constraints. As a result, retrieval may be dominated by evidence that is semantically relevant but temporally incorrect, leading to plausible yet untrustworthy answers. To address this limitation, we propose TA-RAG, a time-aware RAG architecture that treats time as a first-class execution constraint across temporal parsing, indexing, retrieval, and grounded response generation. Our method converts multivariate IoT logs into hourly textual evidence, normalizes temporal expressions into executable retrieval plans, restricts retrieval to the relevant temporal partitions, and enforces timestamped citations in the final answer. Evaluated on real smart home sensor logs, compared with standard RAG, TA-RAG improves temporal plan correctness from 20.83% to 66.67%, increases retrieval precision and coverage while reducing temporal leakage, and achieves 96.96% citation compliance with only 3.87% unsupported claims. These results show that plan-constrained retrieval is an effective design principle for trustworthy natural-language analytics.