CoMuRAG: Condition-Mutated Hard Negatives for Safety-Sensitive Korean Building-Code Retrieval

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Retrieval over building structural design codes is safety-sensitive: two queries that are nearly identical in surface form can require different table cells when a single condition changes, and a plausible value may come from the wrong standard. We formalize this problem as condition-confusion retrieval over Korean Design Standard (KDS) documents and measure it with the Condition-Confusion Rate (CCR), the fraction of queries for which a condition-mutated distractor outranks the gold node. To reduce CCR, CoMuRAG indexes KDS PDFs into dual-view nodes and mines five balanced, document-grounded families of hard negatives (axis-neighbor, exception-flip, unit-disguise, region-confusion, and cross-document) to fine-tune a cross-encoder reranker. On an eight-document benchmark built under a leakage-free, paraphrase-hardened protocol (n = 431), the naturalized reranker improves CCR over a strong multilingual dense baseline from 0.148 to 0.108 ± 0.003 across five seeds (McNemar p < 0.05 in every seed), with gains concentrated in unit, cross-document, and exception confusions. On an easier single-document benchmark dense leaves little headroom (CCR = 0.071) and reranking does not help, and an ablation indicates that unit-disguise is the strongest training signal whereas axis-neighbor is harmful. Condition-aware reranking is therefore most useful when the benchmark exposes sufficient condition-confusion headroom.