RAGStress: A controlled benchmark for evaluating retrieval-augmented generation under knowledge-base degradation
Retrieval-Augmented Generation (RAG) is typically evaluated under the implicit assumption that the underlying knowledge base (KB) is clean, leaving the behaviour of RAG systems under realistic KB degradation poorly characterised. We introduce RAGStress, a controlled evaluation benchmark for stress-testing RAG systems under systematic KB corruption. The benchmark pairs four naturalistic corruption…
We haven't written up this one. arXiv cs.AI has the full story — the link below goes straight to it.