Vocabularies

Curated GuardVocabulary packs for the highest-traffic languages, and a forLocale lookup.

val cache = SemanticCache(embedder, guards = MatchGuards.standard(Vocabularies.ITALIAN))
// or, equivalently, MatchGuards.standard(Locale.ITALIAN)

The bar these packs are held to

The same one as English: a marker earns its place only if adding it will not reject a genuine paraphrase, and two units alias only when they are spellings of the same quantity. The packs are deliberately conservative — an incomplete marker set makes the guards miss a near-miss (lower recall), which costs one model call; it never turns into a false hit. Grow a pack from your own traffic when you need more, and measure the additions the way the English list was measured.

Each pack is validated against a small language-specific near-miss corpus (see the guard tests): the near-misses are caught and the paraphrases are kept. That is the floor, not a claim of parity with the heavily-tuned English list.