Your AI Is Doing Too Much Work — New Fix Makes It Faster

When you ask an AI assistant a question, it often searches a pile of documents first and then writes an answer from what it found. That two-step approach is called RAG, short for "letting the AI look things up." Right now, most systems do it the same way every time: same searching tool, same brainpower, whether you asked who won a game or something that requires piecing together five separate facts. That's like sending a delivery truck to pick up a single envelope.

A research team from India and the UK studied what happens when you mix and match those two steps instead. Their finding was counterintuitive: spending more on the searching step usually helps more than spending more on the writing step, but only up to a point. Past that, extra effort either does nothing or actually makes things worse. Harder isn't automatically better.

So they built DRAG, a system that guesses how difficult your question is and picks the right combination on the fly. It has two versions. The first, DRAG-QPP, needs no extra training at all — it uses quick signals about the question and the documents to route the request, and it matched the quality of standard setups while cutting the time each answer takes. The second version fine-tunes an AI to predict the best pairing and beat the standard setups on quality, across three different AI model families and four question-answering test sets.

The catch: this is a research paper, not a product you can use today. It's also fiddly engineering — companies would need to wire it into their own systems. But the direction is clear. If AI providers can answer the same questions with less computing muscle, that means faster responses for you and lower bills for them, which usually trickles down to cheaper or free AI features.