Embedding#

Embedding exposes Apple's built-in word and sentence embeddings: dense vectors used for similarity, nearest-neighbor search, and distance computation.

There is no public constructor — get an instance from one of the static factories.

#Static factories

#Embedding.wordEmbedding(language, revision?): Embedding | null

The built-in word embedding for language. Returns null if no embedding is available on this OS for that language.

#Embedding.sentenceEmbedding(language, revision?): Embedding | null

The built-in sentence embedding for language (iOS 14+). Returns null on iOS 13 or when no sentence embedding is available.

const wordEmb = NaturalLanguage.Embedding.wordEmbedding("en")
const sentEmb = NaturalLanguage.Embedding.sentenceEmbedding("en")

#Properties

#Lookup methods

#contains(token: string): boolean

Whether token is in the embedding's vocabulary.

#vector(token: string): number[] | null

Vector for token (length = dimension), or null if the token isn't in vocabulary.

#distance(first, second, type?): number | null

Cosine distance between two tokens ([0, 2]; smaller is closer). Returns null if either token isn't in vocabulary — use contains to disambiguate "no signal" from "very far apart".

#neighbors(token, maximumCount, type?): { token, distance }[]

Up to maximumCount nearest neighbors of token, sorted by ascending distance. Returns [] if token isn't in vocabulary.

#Examples

#Word similarity

const emb = NaturalLanguage.Embedding.wordEmbedding("en")
if (emb) {
  console.log(emb.distance("cat", "kitten"))     // small — semantically close
  console.log(emb.distance("cat", "airplane"))   // larger
}

#Nearest neighbors

const emb = NaturalLanguage.Embedding.wordEmbedding("en")
if (emb) {
  const nbs = emb.neighbors("happy", 5)
  for (const { token, distance } of nbs) {
    console.log(`${token}\t${distance.toFixed(3)}`)
  }
  // joyful  0.341
  // glad    0.402
  // ...
}

#Raw vector

const v = emb?.vector("hello")
if (v) {
  // dot product with another vector ...
}

#Notes

• wordEmbedding("en") is the most reliably available embedding. Other languages may return null depending on OS revision.
• distanceType currently only supports "cosine"; the parameter is reserved for future metrics.