New ARGUS method patches critical RAG blind spots, boosting accuracy by 4.5 points

A new paper reveals a critical flaw in Retrieval-Augmented Generation (RAG) systems: neural retrievers have 'blind spots' where they fail to find relevant information. The researchers introduce ARGUS, a pipeline that detects these gaps using a new Retrieval Probability Score (RPS) and remedies them via targeted document augmentation. ARGUS improved retrieval performance by an average of +3.4 nDCG@5 and +4.5 nDCG@10 across standard models like Contriever and ReasonIR.