Question 24
Domain 6: Human-AI Interaction and OversightYour RAG agent retrieves poorly when users ask multi-part or pronoun-heavy questions in a conversation ('how about that other one I mentioned earlier?'). Select TWO query-rewriting techniques that improve retrieval quality before the search step.
Correct answer: AE
Explanation
Coreference resolution turns pronoun-heavy follow-ups like “that other one I mentioned earlier” into a self-contained query by using prior conversation context. This improves retrieval because the search step receives explicit terms instead of ambiguous references, which aligns with query rewriting before retrieval.
Why each option is right or wrong
A. Rewrite the user's pronoun-heavy follow-up into a self-contained query using prior conversation context (coreference resolution), then issue the rewritten query to the retriever.
Coreference resolution is the relevant pre-retrieval rewrite here: the follow-up “that other one I mentioned earlier” must be expanded with antecedents from the dialogue so the search string contains explicit entities and attributes rather than unresolved pronouns. In retrieval pipelines, this is done before the vector/BM25 lookup, because the retriever cannot reliably match ambiguous references; rewriting the query into a standalone form materially improves recall and precision for multi-turn, multi-part questions.
B. Increase the embedding dimensionality so the retriever has more 'room' to disambiguate pronouns.
C. Disable retrieval on follow-up turns and rely on the LLM's parametric memory for everything except the first turn.
D. Pass the raw user message verbatim to the retriever, since modern embedding models are trained on conversational text and handle pronouns automatically.
E. Decompose multi-part questions into sub-questions, retrieve for each independently, and merge or rerank the combined candidate set before passing it to the LLM.