Corfu, Greece · 12 to 15 July 2026 · Session: Artificial Intelligence in AEC
Semantic Mapping of Request for Information (RFI) to BIM Elements with Large Language Models
Limon Paul Joy1, Farzaneh Zarei, PhD1, Prof. Mazdak Nik-Bakht, PhD, PEng1
1Gina Cody School of Engineering and Computer Science, Concordia University, Montréal, Canada
Abstract
Requests for Information (RFIs) remain a critical yet poorly integrated component of construction workflows, typically managed as unstructured text disconnected from Building Information Modeling (BIM) data. Existing studies primarily focus on reducing RFI frequency or response time, with limited focus on linking RFI content to specific BIM elements. This study presents a data-driven framework that combines large language model (LLM)-based intent extraction with BIM metadata reasoning through a staged matching pipeline. By separating semantic cue extraction from rule-based element selection, the framework enhances interpretability and robustness. Validation on experimental and real-world datasets demonstrates reliable element and family-level retrieval.
Keywords: Large language models, Building information modelling, Request for information, Automation in construction
1Introduction
A request for information (RFI) is a formal question about an element the writer assumes the reader can identify: a door, a parapet, a window type on a given level. The building information model (BIM) holds that element as a typed record, yet nothing in the project data connects the two. Answering an RFI therefore means searching drawings and specifications by hand; an estimate cited in the paper puts this at 17 person-hours per RFI, with RFI-related delays accounting for up to 10% of project duration.
2Framework
The prototype is a Revit add-in bridged to a Python service. A large language model reads each RFI and returns a strictly formatted set of intent cues: discipline, primary object, keywords, spatial references and dimensions. It is not permitted to name BIM identifiers, Revit categories or element instances.
Selection then runs through five ordered stages against the model's own metadata, each contributing an interpretable score: direct identifier, spatial context, system context, a lexical prescore and embedding similarity. A person confirms the result before the link is written back (Figure 1).
3Validation
On an experimental model, 8 of 10 RFIs were linked to the correct element at top-1, and 90% including family or category matches. On a real project, 74 of 524 issued RFIs carried enough text to attempt a link. Accuracy with GPT-4.1 was 60.27% with and without intent extraction.
4Discussion
Performance was driven primarily by the rule-based matching and the quality of the BIM metadata. Failures came from RFIs referring to rooms without instance-level room associations, and from context held in marked-up drawing attachments rather than in the RFI text.
