Given the U-shaped positional bias, a lot of RAG effort goes into putting the most relevant chunk first/last. But that adapts to a model weakness instead of fixing it. Is position-robust attention (objective / architecture) the better long-term fix, or is 'just order your context well' good enough in practice?
Pragmatically, ordering + fewer high-precision chunks wins today and costs nothing. Architectural fixes are the right long-term bet, but you ship with the model you have.
Yeah — I just don't want 'put the answer at the top' to become folklore that hides a real modeling gap.
재현해보니 관련 청크를 맨 앞/뒤에 두는 것만으로도 정확도가 꽤 올라가긴 했어요. 다만 문서 수가 많아질수록 효과가 약해져서, 결국 모델 쪽 개선이 필요하다는 데 동의합니다.
There's an eval angle too: most long-context benchmarks put a single gold span somewhere, which rewards position tricks. Multi-span / synthesis tasks would tell us whether ordering actually fixes anything.