Google researchers showed that five documents echoing a targeted query can reduce retrieval quality by 40.2% to 61.5% in retrieval-augmented generation (RAG) systems. The tests, detailed in an arXiv preprint dated November 3, 2026, covered 623 queries drawn from the SciFact and NFCorpus benchmarks. Both the keyword-based BM25 engine and the dense BGE-small retriever proved vulnerable because query repetition alone could elevate falsified documents. Overlap-based filters blocked some attacks but also reduced the quality of legitimate results, while Maximal Marginal Relevance (MMR) provided little or no protection. The study measures retrieval failures only and does not establish that answer quality would decline by the same percentages.
Source: Read the original article
