最大边际相关性测验

当您参加测验时，就会出现答案和解释。

1. MMR 旨在解决的主要问题是什么？
   • 大型数据库中查询延迟较慢
   • 用户查询中的拼写错误
   • 纯粹相关性排名的冗余、近乎重复的结果
   • 缺乏嵌入的训练数据
2. 在MMR公式中，接近1的lambda值强调什么？
   • 与查询的纯粹相关性
   • 结果之间的最大多样性
   • 所有候选人的权重相同
   • 随机选择
3. MMR 如何选择结果？
   • 通过对单个分数进行排序来一次性完成所有操作
   • 随机使用固定种子
   • 按文件的字母顺序排列
   • 贪婪地，一次一项，与已选择的项目进行比较
4. 基于嵌入的 MMR 实现如何衡量查询相关性和文档冗余度？
   • Using only a document's filename length
   • Using a similarity measure such as cosine, without MMR requiring that one measure universally
   • Using a mandatory audio-transcription score for all text
   • Using no relevance or inter-document comparison at all
5. 为什么 MMR 在检索增强生成 (RAG) 中特别有用？
   • 它提供补充的、非冗余的证据块
   • 它压缩模型的权重
   • 它自动微调语言模型
   • 它对检索到的文档进行加密
6. If MMR caches each candidate's maximum redundancy and updates it against each new selection, roughly how many similarities are needed for k picks from n candidates?
   • Always constant time independent of k and n
   • Exactly n factorial for every implementation
   • About k times n, ignoring vector dimension and initial setup
   • No document comparisons after the first query score
7. With lambda equal to 0.5, relevance 0.8, and maximum selected-item similarity 0.9, what is a candidate's MMR scoring term?
   • 0.85, from averaging the two similarities
   • Minus 0.05, from 0.5 times 0.8 minus 0.5 times 0.9
   • 0.72, from multiplying relevance and redundancy
   • 1.7, from adding the two similarities
8. Can an MMR reranker select a relevant document that the upstream retrieval candidate pool omitted?
   • Yes; diversity scoring invents missing source documents
   • Yes; a lower lambda automatically searches the whole corpus
   • Yes; reranking changes the document's original publication date
   • No; it chooses among the supplied candidates unless the upstream pool is expanded