A city grants portal RAG misses exact NOFO identifiers that applicants type precisely, even when semantic search finds related topics. What should GenAIOps implement?

Hybrid BM25-plus-vector fusion so exact NOFO identifiers complement semantic matches is what GenAIOps should implement when a city grants portal RAG misses exact NOFO IDs applicants type precisely while semantic search still finds related topics. NOFO identifiers behave like serial numbers: keyword methods such as BM25 excel at exact and near-exact token matches, while vectors cover paraphrased opportunity descriptions. Fusion retrieves the precise opportunity applicants named and still supports topical grant browsing. A BM25 lexical index where exact NOFO identifiers supersede semantic matches ranking catches the exact ID but sacrifices the topical coverage the stem says already works, because replacing vectors with lexical-only ranking breaks paraphrased description queries; identifiers should complement semantic matches, not supersede them. A synonym expansion index that expands exact NOFO identifiers into related grant topics dissolves identifier precision into fuzzy topical neighbors and can bury the exact NOFO under related but wrong grants, because it treats a serial number as expandable meaning instead of preserving an exact lexical path. A larger embedding index encoding exact NOFO identifiers as semantic vectors for ranking still treats serial-number tokens as meaning and lacks exact-token precision for hyphens and agency prefixes, so near-identical IDs can outrank the exact identifier without a keyword path. Exam caveat: identifier-centric municipal RAG (NOFOs, case numbers, ISBNs, parcel IDs) nearly always points to hybrid keyword-plus-vector designs. Do not answer with lexical-only supersede or semantic-only encoding of IDs when the stem reports topical success but exact-ID failure and needs both to coexist. Operational check: after enabling BM25-vector fusion, measure exact-NOFO hit rate on identifier queries and topical relevance on description queries, normalize ID tokenization (hyphens, agency prefixes), and confirm fusion scoring does not bury exact BM25 hits under soft semantic neighbors.