The weak link in RAG may be the retriever, not the generator. This paper proposes finding evidence by navigating documents instead of searching an index.
Title: Retriever-Free Retrieval: Navigating Documents with Decision Models
URL:
DecisionRAG organizes the corpus as a tree and lets a decision model (Jev) judge many candidates at once, expanding only the top branches. No task-specific training and no first-stage retriever.
Three highlights:
๐งญ Shared context makes scores comparable
Candidates are judged together, so a plausible one scores lower when a clearly better one is present. On MultiDoc2Dial, scoring passages together instead of page by page lifts R
@1 from 36.7% to 44.4%.
๐ฒ Navigation that scales logarithmically
With a branching factor of 100, 10^9 passages need about five stages, roughly one second. On Natural Questions, keeping about 15% of clusters matches scoring all pages at about one-fifth of the cost.
โก On par or better than GPT-5 reranking, faster and cheaper
Natural Questions Hit
@1 is 65.4 vs. 60.8 for GPT-5 reranking. Latency is 0.68-1.76 s vs. 11.8-15.0 s. Pruning is irreversible, though, and quality depends on the hierarchy.
It reframes retrieval as exploration rather than reranking.
#
RAG# #
InformationRetrieval#