Graph-mediated LLM exploration for IFC question answeringBhatti, Haris Waheed (Avtor) Turk, Žiga (Mentor) master thesiscivil engineeringIndustry Foundation Classes (IFC)Building Information Modelling (BIM)knowledge graphLinked Building Datalarge language modelsadaptive explorationnatural-language question answeringIndustry Foundation Classes (IFC), the standard format for exchanging building information, are hard to query without specialist tooling, so most people working in construction cannot get information out of their own building models. This thesis asks whether mediating IFC through a knowledge graph and exploring it adaptively can compete with exploring raw IFC directly. We built ifctoneo4j, which converts IFC into a Neo4j property graph following the Linked Building Data ontologies, and BIMGraphExplorer, which lets an LLM query that graph in Cypher through Model Context Protocol tools, and evaluated both under matched conditions: same model, same judge, 367 questions across 23 IFC-Bench v2 datasets. The direct-IFC baseline wins by 10.3 percentage points (82.3 % versus 72.0 %) with equal transparency, though the graphmediated system wins outright on five of the twenty-three datasets, by as much as 24 points on one; most of the remaining gap comes from fixable problems in how the graph was built, with the one real ceiling on quantities Cypher cannot compute. This matters because building data can live in an open, shared graph format and still answer questions nearly as well as reading a proprietary file directly, which is useful for anyone building natural-language tools over construction data or working in Linked Building Data. What is left is mostly engineering work: improving how files are converted, and giving the system a way to compute a value instead of only looking one up, so it can also handle geometry and material questions the graph alone cannot answer.[H. W. Bhatti]20262026-07-25 08:45:10Magistrsko delo/naloga185169UDK: 004.42:69(043.2)VisID: 180569COBISS_ID: 286173955sl