RAG research is consolidating around the view that retrieval—not generation, context length, or raw scanning—is the binding constraint, elevating reranking and evidence coverage as independent optimization stages over brute-force context expansion.
Evidence
- ‘The Bottleneck Is Retrieval, Not Reading’ and ‘Scan or Index?’ both show exhaustive scanning or long-context reading fails to beat hybrid retrieval with reranking on evidence coverage and answer quality.
- Reranking is repeatedly shown to be a separate, high-leverage stage: dedicated reranker fine-tunes beat decision-tuned general models on nDCG@10, and staged hard-negative mining outperforms simply enlarging candidate pools.
- Failure modes reinforce this—embedders are ‘metrically affect-blind’, chunking destroys document context, and long-context windows introduce poisoning/distraction rather than eliminating RAG.
Implications
- Teams seeing diminishing returns from bigger embedding pools or longer context should reallocate effort to reranker training and content-aware chunking.
- RAG evaluation should measure evidence coverage and multi-fact retrieval, not just end-to-end answer scores.
Concepts
RAG Reranking Embeddings Long Context Chunking RAG Evaluation Context Engineering
Confidence
high