Input quality is almost always the actual bottleneck. Teams spend months tuning retrieval while feeding HTML boilerplate into their vector stores.

Plenty of folks obsess over retrieval but daily quries are word salad and the input is still garbage.