A retriever can return three relevant-looking results and still miss part of the user’s intent. For a “warm winter jacket” query, three similar warm-jacket results may score well on Precision@3, while a set that also includes a fleece jacket may be more useful despite that item’s marginal relevance. α-nDCG is an intent-aware ranking metric for evaluating coverage of different query facets rather than relevance alone. section 8.6
Ordinary NDCG rewards relevant results appearing earlier in a ranking. α-nDCG adds an explicit concern for whether results cover different subtopics, so repeated coverage does not substitute for breadth. That makes it a candidate when evaluating diverse results from RAG-Fusion. Precision@k remains useful for counting relevant results, but can favor a narrow set. section 8.6 section 14.1.1
Use α-nDCG to ask whether a ranking surfaces distinct relevant facets, not merely more hits on one facet. No α-nDCG calculation or ground-truth construction procedure is specified here; choosing the metric is distinct from implementing a full RAG evaluation. section 8.6
