Retrieval-augmented generation (RAG) has become the standard way to ground large language models in external knowledge. But a new arXiv paper, "The Bottleneck Is Retrieval, Not Reading," argues that the approach's core weakness lies in how it retrieves, not in how the model reads.
The authors diagnose RAG's similarity-driven, chunked pipeline as having structural blind spots: it cannot anticipate which passages a long-context question will actually need. Because retrieval happens before generation, missing evidence at that stage cannot be recovered later, no matter how strong the model's comprehension is.
The paper's title captures its central claim: for evidence-based generation in long-context QA, the bottleneck is retrieval. This shifts attention from improving readers to rethinking how evidence is selected and presented in the first place.