FlagThis

Abstract

Vector hubness—the tendency of a few points to be nearest neighbours of disproportionately many queries in high-dimensional space—is an attack surface for retrieval-augmented generation (RAG): an adversary injects one crafted document that is retrieved for many unrelated queries, turning a single record into a broad poisoning vector. The prevailing defence is detection: periodically scan the index, estimate reverse-k NN influence, and remove outliers, which leaves an exposure window and requires repeated corpus-wide rescans. We study admission-time control—scoring each document's hub behaviour against sentinel queries and refusing it before it becomes retrievable—and the system's question of whether the control can be maintained incrementally rather than by rescans.