Answers are grounded in this paper and cite their sources.
Knowledge bases for primary power equipment must retrieve concise, verifiable evidence from long standards, fault cases, and maintenance records. We evaluate six retrieval-unit strategies, three retrievers, and two Chinese embedding encoders using 149 documents (1.73 million characters) and 48 author-constructed questions, with 20 for development and 28 held out for testing. At the matched 500-character/100-overlap setting, recursive splitting preserves 84.5% of numbered standards clauses, compared with 78.6% for sliding-window, 72.0% for paragraph, and 64.9% for semantic chunking; clause-aware splitting reaches 88.3%. Recursive splitting with dense retrieval records Hit@10 of 0.571 and evidence coverage@10 of 0.542; hybrid dense–BM25 fusion raises these to 0.714 and 0.685. Sliding-window dense retrieval records 0.786 and 0.685 on the same measures. Recursive, sliding-window, paragraph, and semantic chunks each contain all 112 annotated evidence spans, whose lengths range from 11 to 90 characters (median 36); clause-aware chunking contains 97.3%. The results characterize the structural-preservation and retrieval profiles of the tested strategies and show that hybrid fusion improves top-10 evidence recovery with recursive chunks.
No takes yet. Share an insight, caveat, or question.
Cheng et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: