Ranking Free RAG: Replacing Re-ranking with Selection in RAG for Sensitive Domains

arXiv:2505.16014v5 Announce Type: replace

Abstract:

Retrieval-Augmented Generation (RAG) systems operating within sensitive sectors face the dual challenge of delivering transparent evidence selection and maintaining strong defenses against data poisoning. However, existing methods typically depend on black-box similarity retrieval with fixed top-k limits, which lack explanatory power and remain susceptible to adversarial attacks.

To address these limitations, the METEORA framework introduces a rationale-driven selection process that eliminates traditional re-ranking. This approach utilizes three core components: a Large Language Model (LLM) fine-tuned via Direct Preference Optimization (DPO) to produce explicit retrieval rationales; an Evidence Chunk Selection Engine (ECSE) that leverages these rationales alongside statistical elbow detection to determine adaptive cutoffs; and a Verifier LLM that filters out poisoned data based on the same rationales.

Evaluations across six datasets demonstrate METEORA’s superior performance, recording a 13.41% increase in recall, a 21.05% boost in precision (excluding expansion), and an 80% reduction in the volume of evidence required. Furthermore, the system achieved a 33.34% gain in answer accuracy and a 4.4-fold improvement in adversarial robustness. Human assessments validated the system’s genuine interpretability, showing an 86% agreement with ground truth and a confidence score of 3.64 out of 5. These results indicate that interpretability, efficiency, and robustness can be achieved synergistically rather than as competing goals. The source code is publicly accessible at https://github.com/YashSaxena21/METEORA.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC