Back to AI Research

AI Research

DecisionRAG navigates document structure with token-free branch decisions

Key Takeaways

  • A retrieval approach jointly evaluates candidate branches rather than relying on a first-stage similarity retriever.
  • [Retriever-Free Retrieval](https://arxiv.org/abs/2610.35868) introduces DecisionRAG, a system that uses pretrained language-model capabilities to navigate a document collection.
  • The authors address a limitation of retrieval-augmented generation: an answer generator cannot use evidence that its retrieval stage fails to find.
  • DecisionRAG traverses the corpus’s existing structure, moving from broad topics towards individual passages.
  • At each level, a decision model jointly evaluates candidate branches in a single forward pass without generating output tokens.

Retriever-Free Retrieval introduces DecisionRAG, a system that uses pretrained language-model capabilities to navigate a document collection. The authors address a limitation of retrieval-augmented generation: an answer generator cannot use evidence that its retrieval stage fails to find.

Choose branches in a shared context

DecisionRAG traverses the corpus’s existing structure, moving from broad topics towards individual passages. At each level, a decision model jointly evaluates candidate branches in a single forward pass without generating output tokens.
The authors argue that evaluating candidates together makes relevance estimates comparable. This differs from using a fixed embedding similarity as the first filter, then asking a separate model to rerank only the survivors. The proposed system does not use a first-stage retriever or task-specific training in the reported setup.

Reported retrieval results

The paper reports a 13.8-point R@1 improvement on MultiDoc2Dial and a 17.4-point Hit@1 improvement on Natural Questions against dense retrieval with an 8B reranker. Those figures belong to different named retrieval metrics and should retain their benchmark labels.
Across four benchmarks, the authors report retrieval times of 0.7 to 1.8 seconds per query, including evaluation on 2.68 million Wikipedia paragraphs. These are retrieval measurements in the paper’s environment, rather than end-to-end answer latency for an arbitrary application.

Inspect the collection structure

The method makes document organisation part of the retrieval process. A team considering it should examine whether its corpus has useful branches and whether a relevant passage can be reached reliably through them.
A practical comparison would keep the corpus, queries and relevance judgements fixed, then measure retrieval quality and latency together. The reported gains motivate that test; they do not establish that structural navigation will outperform every dense-search configuration or transfer unchanged to a differently organised collection.

Comments