Retrieval-aware RAG chunking — interactive demo

Two chunking strategies side by side on a 1000-document Wikipedia bench with 1032 Natural Questions, three ranking arms sharing one BGE top-20 pool.

What the study found: chunk size dominates recall (size effect ≈ 18× the largest chunking-method effect); fixed, BiLSTM and Transformer chunking tie at a matched size. The one intervention that moved the sweet spot was fine-tuning the cross-encoder reranker (+0.107 R@1 in-domain) — try the third arm and watch chunks climb the ranking.

Code, data and full write-up