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