An Agent-Oriented Framework for CityGML Querying: Retrieval and Deterministic Workflow Evaluation

Querying CityGML building models requires access to structured semantic information and a clear distinction between retrieved evidence and answer correctness. This study presents an agent-oriented framework combining metadata extraction, deterministic query coordination, evidence validation, and bounded recovery. Evaluation uses four static CityGML 2.0 building models containing LoD4 geometry and 4,181 processed metadata entities. A standalone retrieval experiment compares BM25 and dense E5 using 80 queries, while a separate 32-query benchmark compares direct metadata processing, a workflow without validation and recovery, and the full workflow against independently reconstructed reference answers. BM25 achieved overall Precision@5 of 0.5225, Recall@5 of 0.5325, and MRR@10 of 0.9150, outperforming dense E5 under the evaluated representation. Median retrieval latency was 4.80 ms for BM25 and 68.10 ms for dense E5. Analysis by query type and reference-set size distinguished ranking performance from the attainable Recall@5 ceiling; document truncation also removed query-defining information from some inputs. Direct processing and the ablated workflow each produced 32 correct final outcomes, compared with 26 for the full workflow, whose validation and recovery increased latency without improving correctness on this workload. Metadata preprocessing required approximately 1.47–15.15 s. These results inform the design of CityGML query workflows by distinguishing retrieval coverage, structured-result correctness, and final-answer reliability.