Spend 80% of RAG effort on retrieval: contextual chunks + cross-encoder rerank, recursive-512 as the default split

Digital Applied·medium signal

The 2026 consensus is that retrieval, not generation, is the bottleneck. Prepend a 1–2 sentence context summary to each chunk (Anthropic's Contextual Retrieval cuts top-20 failures up to 67% with reranking), run a two-stage pipeline that pulls 50–100 bi-encoder candidates then rescoring them with a cross-encoder down to the top 3–10, and start from recursive 512-token splitting, which topped a February 2026 seven-strategy benchmark. Chunking choice alone can swing recall up to 9% on the same corpus, so treat it as a tuned parameter, not a default you set once.