引導式測驗 · 中等 等級

檢索指標:Recall@K、MRR 與 nDCG 測驗

根據排名目標和相關性標籤選擇 Recall@K、MRR 和 nDCG,並了解哪些檢索分數未測量。

問題 1 的 8

What does Recall@K measure?

預覽每個問題8 問題

當您參加測驗時,就會出現答案和解釋。

  1. What does Recall@K measure?
    • The rank of only the first relevant item
    • The fraction of known relevant items retrieved within the top K results
    • The generation quality of a final answer
    • The average latency of a search
  2. Which aspect does MRR emphasize?
    • The total number of relevant results after rank K
    • Graded relevance of every result
    • Whether a generated answer cites a source
    • How high the first relevant result appears in the ranking
  3. What type of judgment can nDCG use?
    • Graded relevance, with lower-ranked items discounted
    • Only a model’s output token count
    • Whether a request timed out
    • Only exact-match labels for the generated answer
  4. Why might MRR be suitable when users need one result quickly?
    • It rewards retrieving every relevant item equally
    • It measures whether all documents were indexed
    • It rewards placing the first relevant item near the top
    • It evaluates answer factuality directly
  5. What does Recall@K fail to distinguish by itself?
    • The number K used
    • Different orderings of relevant items that are both within the top K
    • Whether the relevant set is known
    • The query text
  6. Why can incomplete relevance judgments distort retrieval metrics?
    • An actually relevant but unjudged result may be treated as nonrelevant
    • Metrics automatically fill in all missing labels
    • Every query has exactly one relevant result
    • Judgments affect only model latency
  7. What should be reported when comparing nDCG implementations?
    • Only the retrieval model name
    • The number of GPU cores
    • Only the prompt template
    • The gain and discount conventions, cutoff, and aggregation
  8. Does a high retrieval Recall@K prove that a RAG answer is correct?
    • Yes, if K is at least ten
    • Yes, because relevant documents guarantee faithful synthesis
    • No; it measures retrieval coverage, not final answer correctness
    • No, because retrieval cannot be evaluated