What really breaks RAG systems: long context or hard competitors?

RAG systems retrieve relevant passages but still fail at reading comprehension. This work disentangles two failure modes: context length versus semantic competition from misleading passages. By swapping difficult distractors with weaker alternatives while holding passage count and length fixed, the authors recover 4–6 exact-match points on SQuAD with Phi-2 and Qwen2.5-1.5B, plus larger gains in answer inclusion and F1. Competition effects are real but separate from context bloat.