Agentic search for BIM information extraction: Systematic evaluation

Although Building Information Models (BIM) help structure building data, retrieving specific information from them remains difficult for most project stakeholders. Natural-language (NL) interfaces can help mitigate this issue. Conventional approaches translate NL queries into structured queries that assume a fixed data organization, which BIM heterogeneity routinely invalidates. This paper addresses this limitation through agentic search, an LLM-based agent that iteratively retrieves information from BIM models while discovering their structure at runtime rather than relying on predefined assumptions. The approach is evaluated using ifc-bench v2, an open-source benchmark released with this work, comprising 1027 question–answer pairs across 37 IFC models. A 3 × 2 × 2 factorial experiment reveals that agentic search outperforms the static baseline in every configuration tested (+29.5 to +48.5pp). A detailed analysis uncovers the approach’s inner workings and remaining pitfalls, while a regression analysis ranks the factors driving accuracy: the search paradigm, followed by BIM-model quality and backbone capability.