DecisionRAG
A retriever-free retrieval-augmented generation method that uses a decision model to jointly evaluate multiple candidate branches in a single forward pass, navigating the hierarchical structure of a corpus without task-specific training or a first-stage retriever.
Timeline (last 90 days)
DecisionRAG proposes a retriever-free retrieval paradigm that does not rely on a first-stage retriever and does not generate any tokens, but instead uses a decision model to jointly evaluate multiple candidate branches in a single forward pass.
DecisionRAG navigates hierarchically along the inherent structure of the corpus, descending level by level from broad topics down to individual paragraphs, deciding at each level which branches are worth exploring further.
The core design of DecisionRAG is to place multiple candidate branches in a shared context for joint judgment, making relevance estimates relatively comparable, which differs from traditional pointwise scoring reranking.
On the MultiDoc2Dial dataset, DecisionRAG, without task-specific training and without a first-stage retriever, achieves 13.8 higher R@1 points than dense retrieval using an 8B reranker.
DecisionRAG's evaluation covers four benchmarks, including a corpus containing 2.68 million Wikipedia paragraphs.
DecisionRAG's retrieval takes only 0.7 to 1.8 seconds per query.
The significance of DecisionRAG points to an architectural-level reflection: whether retrieval and generation must be separated; when the language model is powerful enough, handing retrieval decisions to the model itself may be more flexible than maintaining a separate embedding index.
DecisionRAG still faces questions to be validated, including scalability on large-scale corpora, the compatibility cost with existing indexing systems, and the robustness of the decision model under domain drift.
All Facts (8)
DecisionRAG proposes a retriever-free retrieval paradigm that does not rely on a first-stage retriever and does not generate any tokens, but instead uses a decision model to jointly evaluate multiple candidate branches in a single forward pass.
50%UnverifiedDecisionRAG navigates hierarchically along the inherent structure of the corpus, descending level by level from broad topics down to individual paragraphs, deciding at each level which branches are worth exploring further.
50%UnverifiedThe core design of DecisionRAG is to place multiple candidate branches in a shared context for joint judgment, making relevance estimates relatively comparable, which differs from traditional pointwise scoring reranking.
50%UnverifiedOn the MultiDoc2Dial dataset, DecisionRAG, without task-specific training and without a first-stage retriever, achieves 13.8 higher R@1 points than dense retrieval using an 8B reranker.
50%UnverifiedDecisionRAG's evaluation covers four benchmarks, including a corpus containing 2.68 million Wikipedia paragraphs.
50%UnverifiedDecisionRAG's retrieval takes only 0.7 to 1.8 seconds per query.
50%UnverifiedThe significance of DecisionRAG points to an architectural-level reflection: whether retrieval and generation must be separated; when the language model is powerful enough, handing retrieval decisions to the model itself may be more flexible than maintaining a separate embedding index.
50%UnverifiedDecisionRAG still faces questions to be validated, including scalability on large-scale corpora, the compatibility cost with existing indexing systems, and the robustness of the decision model under domain drift.
50%