A county public-health bot must answer resident questions by meaning when wording differs from the published guidance titles. Which search approach should they configure for grounding?

The county public-health bot must answer resident questions by meaning when wording differs from the published guidance titles. Configure semantic search over public-health guidance so meaning-based questions still retrieve the right documents for grounding. Residents ask about “shots for kids before school” rather than exact PDF titles; semantic retrieval bridges that civic language gap.

Keyword search over public-health guidance with synonym maps so varied queries still retrieve still ranks by literal terms expanded only to hand-listed synonyms; a synonym map can link “shots” to “immunization” only if you manually add it, but it does not provide true meaning-based retrieval and does not understand unseen paraphrases, so varied questions can still miss the right guidance. Hybrid search over public-health guidance with BM25 ranking so varied queries still retrieve combines keyword and vector signals but ranks by lexical BM25 scoring—BM25 lets literal term frequency and overlap dominate, so paraphrased wording still outranks true content meaning. Vector search over public-health guidance with TF-IDF vectors so varied queries still retrieve uses the wrong kind of vector: TF-IDF vectors encode term frequency, not meaning, so they still require literal overlap and do not understand that “shots for kids” means “immunization,” so paraphrased questions can still miss the right guidance.

Exam caveat: semantic search (embeddings / vector retrieval) differs from keyword search with synonym maps that only expand listed terms; many production systems use hybrid search, but when the stem stresses meaning despite wording differences, semantic search tuned to the county’s own index is the intended answer—not keyword synonym maps claiming varied-query retrieval, a hybrid with BM25 ranking, or vector search with TF-IDF vectors. Grounding still requires citing retrieved guidance, not free invention.

Operational check: build a paraphrase test set of resident health questions, compare semantic versus keyword hit rates, require answer citations to guidance sections, and escalate clinical edge cases to nurses. Refresh embeddings when guidance updates so the bot never grounds on superseded public-health PDFs.