Викторина по семантическому поиску

Ответы и пояснения появляются по мере прохождения теста.

1. What distinction describes a dense semantic retriever compared with a purely term-scored lexical baseline?
   • Every lexical system can match only identical raw bytes
   • Dense retrieval guarantees a correct answer for every paraphrase
   • It ranks using learned vector relationships rather than only term-based evidence
   • Semantic retrieval must ignore the query
2. How can a dense search setup score a query against a passage?
   • By comparing compatible query and passage vectors using its trained metric
   • Only by sorting passage filenames
   • Only by taking the latest publication date
   • By requiring identical raw text in every result
3. Why might a dense search experiment use an HNSW index?
   • To guarantee exact top results for every setting
   • To prove every neighbor is factually correct
   • To eliminate the need to encode searchable items
   • To reduce search work while accepting a possible nearest-neighbor recall tradeoff
4. What computation distinguishes a cross-encoder reranking stage?
   • Only offline encoding of the passage without its query
   • Joint evaluation of the query with each selected candidate
   • Retraining the whole corpus encoder for every result
   • Removing all candidates before evaluating relevance
5. Что такое «гибридная» поисковая система?
   • Система, которая сочетает оценку семантических векторов с классической оценкой ключевых слов.
   • Система, которая работает как на телефонах, так и на серверах
   • Система, которая ищет только изображения и аудио
   • Система, которая никогда не использует вложения
6. In a RAG system using semantic retrieval, what does the retriever provide to the generator?
   • A guarantee that every final answer is true
   • Newly trained generator weights for each question
   • Selected source passages as context
   • A replacement for all response generation
7. In the original DPR inference procedure, which representations are computed and indexed offline?
   • Vectors for every possible future question
   • Passage vectors for the searchable corpus
   • Final generated answers for every future user
   • Only scores from jointly processing unknown query-passage pairs
8. What relationship does DPR’s training objective encourage for a question and its positive passage relative to negative passages?
   • An identical score for all positives and negatives
   • A lower positive score than every negative
   • Alphabetical ranking independent of encoder outputs
   • A higher dot-product score for the positive passage