Last year I worked with an AI startup, an Oxford spinout. Their product answered research questions through a RAG pipeline. It worked, but every query took around 90 seconds. Long enough that users were bailing before the answer even loaded. The obvious move is to blame the model and go bigger. That wasn't it. The retrieval layer was doing way more work than it needed to on every single query: bloated embeddings, no caching, redundant calls stacking up as the document set grew. I stripped th...