Year: 2026
Type: conference-paper
Abstract: Retrieval-Augmented Generation (RAG) systems depend on non-parametric indices to access external knowledge, yet most retrieval infrastructure assumes a stationary query document distribution after index construction.In dynamic settings involving continual knowledge updates or evolving terminology, this assumption often fails, leading to degraded retrieval performance, while full re-indexing remains computationally expensive.We propose AURORA, a neuro-symbolic framework for adapting retrieval indices under distribution shift by treating index maintenance as a few-shot continual learning problem.AURORA decouples discrete index structure from continuous metric representations, enabling efficient adaptation of neural components while preserving index topology.A lightweight Bayesian routing policy further balances stability and plasticity by dynamically selecting among adaptive neural indices and static fallbacks based on uncertainty estimates.Across dense, learned sparse (SPLADE), and generative (DSI) retrieval settings, AURORA recovers up to +26.9% Recall@10 on novel topics compared to static baselines, while adapting significantly faster than full retraining (28 ms vs. s).To bridge this gap, we introduce AU-RORA (Adaptive Universal Routing with Online Refinement and Adaptation), a neuro-symbolic framework that treats index maintenance as a Few-Shot Continual Learning problem.AURORA decouples the index topology from its metric representation, freezing the graph structure while adapting the quantization codebooks via a meta-learned update rule.Our key insight is that by treating each distribution shift as a "task" in the meta-learning sense, we can leverage gradient-based adaptation without suffering from catastrophic interference.This work focuses on maintaining retrieval effectiveness under distribution shift.AURORA is a modular framework that adapts existing dense, sparse and generative retrievers using lightweight continual learning mechanisms, without modifying underlying encoder architectures or index topology.Our contributions are as follows:• We frame retrieval index maintenance under distribution shift as a few-shot continual learning problem and propose a neuro-symbolic framework that enables efficient adaptation without full re-indexing.• We introduce a meta-learned adaptation mechanism that supports localized embedding updates for novel or drifting vocabulary using a small number of gradient steps.• We propose an uncertainty-aware routing policy that dynamically selects among dense, sparse and generative retrieval modalities to balance robustness and adaptation speed.• We empirically evaluate the framework across dense, learned sparse (SPLADE), and generative (DSI) retrieval settings, demonstrating improved recall and substantially faster adaptation compared to static baselines.
Institutions: Virginia Tech, Columbia University
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W7166874927