<![CDATA[Building Information Modeling (BIM) consolidates heterogeneous building information into a single digital model, yet retrieving meaningful insights from Industry Foundation Classes (IFC) files remains difficult and demands specialised expertise. Recent work has shown that converting IFC into a Linked Building Data (LBD) knowledge graph (KG) and pairing it with Large Language Models (LLMs) enables natural language querying, but three limitations persist: (i) the full ontology cannot be supplied to the LLM because of prompt-token constraints, restricting holistic reasoning; (ii) single-pass SPARQL generation fails on complex queries, which is the dominant error mode; and (iii) implicit spatial concepts that are not explicitly modelled as KG entities, such as rooms, corridors and enclosed areas, are frequently answered incorrectly. We introduce AGORA-BIM, a framework that extends the KG-plus-LLM paradigm with three coordinated modules: a retrieval-augmented ontology-grounding module that injects only the semantically relevant schema fragments into the prompt; an agentic, self-correcting SPARQL-generation module built around an executor–validator–repair loop with verbal reflection; and a spatial-reasoning module that reconstructs implicit building entities from element geometry, adjacency and opening relationships. We describe the architecture, the prompting design, and an evaluation protocol on a real multi-storey office building in Barcelona with an extended question set, comparing AGORA-BIM against the single-pass baseline and across commercial and open-source LLM backbones. The illustrative results reported here indicate that grounding LLM reasoning in retrieved schema, bounded self-correction, and explicit spatial inference together substantially improve correctness on indirect and reasoning-intensive questions, while preserving interpretability. AGORA-BIM thus advances intuitive, scalable and cost-aware natural language access to BIM data.Keywords: Building Information Modeling (BIM), Large Language Models (LLMs), Retrieval-Augmented Generation (RAG), Knowledge Graphs, Agentic AI]]>
Copyrights © 2026
