引導式測驗 · 硬 等級
查詢重寫和多查詢檢索測驗
測試 RAG 的查詢轉換:會話重寫、分解和後退提示、多查詢檢索、倒數排名融合和 HyDE。
問題 1 的 8
為什麼在多輪對話中 query rewriting 尤其重要?
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。
- 為什麼在多輪對話中 query rewriting 尤其重要?
- Follow-up messages often depend on earlier context and are meaningless as standalone searches
- Chat models cannot generate text
- Vector databases reject questions
- Longer queries are always faster
- 哪種技術會將比較問題拆分為針對每個被比較項目的獨立搜尋?
- HyDE
- Reciprocal rank fusion
- Decomposition
- Prompt caching
- 回溯提示(step-back prompting)有什麼作用?
- Deletes the user's query
- Retrieves only the most recent documents
- Reverses the order of search results
- Asks a more general question first to retrieve background principles
- 在 reciprocal rank fusion 中,文件的分數如何計算?
- The average of raw similarity scores
- The sum of 1/(k + rank) across result lists
- The number of words it shares with the query
- Its rank in the first list only
- 為什麼 RRF 在合併向量搜尋與 BM25 結果時很方便?
- It relies on ranks, so scores on different scales need no normalization
- It requires identical scoring functions
- It only works with a single list
- It retrains the embedding model
- What does HyDE embed instead of the user's question?
- The document title
- A random sample of chunks
- A hypothetical answer written by the model
- The user's conversation history
- What is a recommended safeguard against intent drift in multi-query retrieval?
- Generate at least twenty variants
- Remove all named entities from variants
- Skip deduplication
- Include the original query as one of the variants and preserve entities and dates
- Which is a real drawback of generating many query variants?
- It makes RRF impossible
- It adds latency and cost and can pull in irrelevant documents
- It prevents reranking
- It always lowers recall to zero