Yuning Bian and colleagues compare document splitting methods across 120 adjacent Wikipedia revision pairs. A small edit can shift boundaries and force a retrieval system to embed unchanged text again.
- On the 116 pairs all methods process successfully, English fixed windows recompute 57.4% of new-version input. Paragraph-first splitting recomputes 37.9%, and content-defined splitting 42.8%.
- Embeddings are numerical representations used for search. Reuse here requires byte-identical complete input under the same model and a cache of the previous document version.
- Under a 2,048-token retrieval budget, English evidence recall is 96.7%, 96.8% and 95.0%, respectively. That measures supporting-passage retrieval, not final-answer accuracy.
- The two content-aware methods change ranking with length settings. Content-defined splitting fails on four Chinese articles with long whitespace-free spans.
Test update work alongside static retrieval quality. A simple paragraph-first baseline deserves comparison before adopting a more elaborate splitter.
Caveat: Recomputed input is work, not elapsed time or money. Small recall differences do not prove the methods have equivalent answer quality.