Retriever-Free Retrieval: Navigating Documents with Decision Models

TL;DR

DecisionRAG introduces a retriever-free approach that uses decision models to navigate document corpora hierarchically, evaluating candidate branches in a shared context without generating tokens. It outperforms dense retrieval with an 8B reranker by up to 17.4 Hit@1 points on benchmark datasets, achieving fast retrieval in under 2 seconds per query. The method eliminates the need for task-specific training or first-stage retrievers.

Why It Matters

This approach closes the capability gap between retrieval and generation by enabling language models to directly assess relevance across large corpora, reducing reliance on brittle lexical matching and enabling faster, more accurate retrieval without additional training or infrastructure.

Key Findings

• Decision models can jointly evaluate multiple document branches in a single forward pass without token generation.
• Shared context during evaluation enables comparable relevance estimates across candidates, improving retrieval accuracy.
• DecisionRAG achieves state-of-the-art performance on multiple benchmarks without a retriever or task-specific training.