Retriever-Free Retrieval: Navigating Documents with Decision Models
DecisionRAG uses an LLM to navigate document branches without a first-stage retriever. Its reported gains suggest retrieval quality can improve by changing how candidates are searched, not only how they are ranked.
DecisionRAG uses a language model to choose which branches of a corpus to explore, then narrows from broad topics to passages. It generates no tokens during these decisions and requires no task-specific training or first-stage retriever. Across four benchmarks, it ran in **0.7–1.8 seconds** per query, including on **2.68 million** Wikipedia paragraphs.
If your agent depends on retrieval, this points to a possible alternative to lexical or embedding search followed by reranking: use a decision model to guide corpus navigation. The reported comparison beat dense retrieval with an 8B reranker by **13.8 R@1 points** on MultiDoc2Dial and **17.4 Hit@1 points** on Natural Questions.
DecisionRAG uses a language model to choose which branches of a corpus to explore, then narrows from broad topics to passages. It generates no tokens during these decisions and requires no task-specific training or first-stage retriever. Across four benchmarks, it ran in **0.7–1.8 seconds** per query, including on **2.68 million** Wikipedia paragraphs. If your agent depends on retrieval, this points to a possible alternative to lexical or embedding search followed by reranking: use a decision model to guide corpus navigation. The reported comparison beat dense retrieval with an 8B reranker by **13.8 R@1 points** on MultiDoc2Dial and **17.4 Hit@1 points** on Natural Questions. These are paper results, and the material does not describe a production implementation or its deployment costs. It remains open how well this approach transfers to other corpora and workloads.
DecisionRAG presents corpus navigation as an alternative retrieval route: a model selects branches and passages without a first-stage retriever. Its reported benchmark results make it a candidate to compare with existing pipelines, while the supplied work on chunking, memory, and persistent knowledge shows that retrieval quality and cost remain workload- and architecture-dependent.