A housing authority wants resident policy Q&A grounded in lease embeddings so paraphrased questions still retrieve the right clauses. Which search capability should they configure?

Vector search over lease embeddings is the capability a housing authority should configure when resident policy Q&A must survive paraphrase. Lease language uses formal section titles and defined terms; residents ask about pets, deposits, notice periods, and repairs in everyday words. Dense embeddings place those paraphrases near the correct clause chunks so retrieval can ground answers without requiring exact lease wording. HNSW ANN lookup over lease metadata fails for paraphrased policy questions because it traverses metadata values rather than lease clause embeddings, so it does not compute semantic similarity for meaning-based retrieval. Semantic reranker rescoring of BM25 matches only reorders already retrieved keyword candidates and does not perform initial vector retrieval over lease embeddings, so it cannot surface the right clause when paraphrased questions share no exact lease terms. kNN scoring over lease keyword index fails for paraphrased policy questions because a keyword index holds exact terms without dense lease embeddings, so there are no vectors to score for meaning-based retrieval. Exam caveat: when the stem stresses paraphrased municipal policy questions against document embeddings, choose vector (semantic) search—not HNSW ANN lookup over lease metadata, semantic reranker rescoring of BM25 matches, or kNN scoring over lease keyword index. Operational check: embed approved lease clause chunks with metadata for property and effective date; evaluate with paraphrased resident questions that omit section numbers; confirm the correct deposit, pet, and notice clauses retrieve before wiring the housing Q&A agent to the index.