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。
- 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
- 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
- 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
- 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
- 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
- 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
- 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
- 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