Skip to content
Trending storyDeveloping

Google paper shows query-copying documents can skew RAG retrieval rankings

1 article1 sourcesince Nov 5Last article 2h ago ·

Overview

AISummary of 1 article

Google researchers report that adding five documents which repeat a user's query can sharply shift retrieval rankings in RAG systems, according to a paper cited by X user Rohan Paul.

Classic retrievers lost 40.2% to 61.5% ranking quality across 623 queries. The attack payload itself barely changed which documents were retrieved, so the query-copying step does most of the work.

The paper also reports that a filter blocking query-echoing documents halved clean ranking quality on NFCorpus, from 0.267 to 0.134.

Written by AI from the articles below · updated Nov 5, 1:53 PM ET

Check the sources:

Latest developmentNov 5, 1:24 PM ET
Google paper finds query-copying documents poison RAG retrieval

Article timeline

The articles in this story. Times are ET.

Nov 5
  1. Rohan PaulX
    Google paper finds query-copying documents poison RAG retrieval

    AIGoogle researchers report that adding five documents which repeat a user's query can shift retrieval rankings sharply in RAG systems, with classic retrievers losing 40.2% to 61.5% ranking quality across 623 queries. The attack payload itself barely changed which documents were retrieved, so the query-copying step does most of the work. A filter that blocks query-echoing documents also halved clean ranking quality on NFCorpus, from 0.267 to 0.134.

    Image from @rohanpaul_ai's post

Heat Trend

Not enough continuous observations to show a trend yet.