Retrieval-augmented generation systems usually run every query through the same fixed retriever and generator setup. New research says that is wasteful, and proposes swapping configurations on the fly instead.
A new paper introduces DRAG, a framework that picks retriever and generator settings per query rather than using one static pipeline for everything. The researchers first show that stronger retrieval tends to help more than throwing extra generation effort at a problem, but both deliver diminishing and sometimes inconsistent returns, so the most complex setup is not always the best one. They test two versions: a training-free variant that uses query performance prediction and perplexity signals to route each query, and a fine-tuned variant that trains an LLM to jointly choose retriever and generator configurations. Tested across three LLM families and four question-answering benchmarks, the training-free version matches strong static baselines while cutting inference latency, and the fine-tuned version beats both static and other adaptive approaches on effectiveness.
This matters because most RAG deployments still treat every question the same, whether it is a simple factoid lookup or a multi-hop reasoning chain that needs several retrieval passes. That means overpaying in compute for easy queries and underserving hard ones. A routing layer that scales effort to actual query difficulty is a more direct fix than the usual approach of just making retrieval or generation individually smarter.
It is a sensible idea dressed in a slightly clunky acronym, and the real test is whether the routing overhead stays cheaper than the compute it saves once this leaves benchmark conditions for messier production traffic.