DecisionRAG replaces RAG retrievers with a top-down decision model

DecisionRAG, described in arXiv:2610.35868, replaces the first-stage retriever in RAG with a decision model that moves from broad topics to documents to passages, scoring candidates in one forward pass without generating tokens.

The paper reports that, with no first-stage retriever and no task-specific training, it beats dense retrieval plus an 8B reranker by 13.8 R@1 points on MultiDoc2Dial and 17.4 Hit@1 points on Natural Questions, searching 2.68 million Wikipedia paragraphs.

A Qwen3-8B judge trained on in-domain data reaches 60.7 R@1 on MultiDoc2Dial, 5.7 points above the best published task-trained retriever, according to the paper. Elvis Saravia reports retrieval takes 0.7 to 1.8 seconds and 0.2 to 2.3 cents per query.