Semantic Deduplication of Queries

String matching catches a repeated question only when somebody types it identically, which almost never happens. Real users ask about parental leave as what is our parental leave entitlement, how many weeks do I get when the baby comes, and the same enquiry in Spanish — three phrasings, one question, three billable calls under conventional caching. Semantic deduplication resolves the phrasings to a single intent, so the governed answer is computed once and served thereafter regardless of wording or language. The effect is not marginal. In organizations of any size the distribution of questions is extremely concentrated, and the difference between counting enquiries and counting distinct questions is usually an order of magnitude.

Centralpoint meters return trips to the model against intent rather than against text, so a question already answered is recognized as already answered however it arrives. Governed answers are served from the local index, which means the thousandth phrasing of a settled question costs nothing and arrives faster than the first. The measurement is the organization's own, held in its environment, rather than a figure supplied by the party being paid.