Structure-aware retrieval-augmented generation for industrial standard documents

Abstract

In standard question scenarios, the closest sentence is not always the correct one. We have found cases where retrieval returns only the definition and fails to include the note, table or formula condition necessary for using the definition. To address this issue, we construct retrieval units centered on structural clauses. Each parsed node keeps a section path and any local structured information, so the retriever does not treat a clause number or a table as disposable layout. Clause-level chunks are retained unless they exceed the 512-character fallback limit. Dense Search handles paraphrasing; inverted index search handles codes and terms; a cross-encoder then re-ranks the fused results. Recall@5 for 86 industrial standards and 280 question-answer pairs increases from 0.7393 to 0.7964, and MRR rises from 0.6084 to 0.6627 over Naive RAG. Ablation results show that removing clause-aware chunking harms answer accuracy the most.