Nivas Hegde

Chunking Is Not the Bottleneck: Why RAG Evaluation Must Look beyond Segmentation Strategy

It is argued that, for QA-style retrieval on general-purpose natural-language corpora, chunking strategy is a low-variance factor, and that RAG benchmarks should report variance decomposition alongside absolute metrics so practitioners can allocate engineering effort to the factors that actually drive performance.