BuildingChat: Enabling User Interactions with General-Purpose Sensing Infrastructure for Buildings
TL;DR
BuildingChat is a conversational system that enables building occupants to query sensor infrastructure directly and achieves 88% query accuracy across six language models, outperforming code generation, in-context reasoning, and ReAct agent baselines by 14-46%.
Abstract
Smart buildings collect extensive sensor data, but occupants rarely benefit from it: data flows to facility management while occupants navigate fragmented dashboards or lack access entirely. We first conducted an 11-participant formative interview study to understand building occupants' unmet needs, finding that user query needs vary widely. Question-answering systems can support these dynamic use cases, but smart buildings present challenges unaddressed by general-purpose chatbots: buildings contain thousands of sensors generating continuous data, queries often require multi-step analysis across locations, and the user may need guidance to develop queries that do not invade other building occupants' privacy.; AB@Based on these insights, we developed BuildingChat, a conversational system that enables building occupants to query sensor infrastructure directly. BuildingChat decomposes queries into composable template operations, enforces tag-based access control with policy-driven time resolution, and iteratively refines queries when information is missing or access is denied. We show that BuildingChat achieves 88% query accuracy across six language models, outperforming code generation, in-context reasoning, and ReAct agent baselines by 14-46%. Unlike these baselines, which leak restricted data in 13-46% of queries despite policy instructions, BuildingChat enforces access control by construction. We also conducted a three-day usability study with 10 occupants of a building instrumented with over 300 sensors. Participants found BuildingChat easy to learn and its responses understandable, rating it 80±8 on the system usability scale.