@glyphp/resolver

Maps a natural-language intent to candidate glyphs, ranked by a pluggable Scorer.

Default — zero dependencies

import { GlyphResolver } from '@glyphp/resolver'

const lexicon = await client.getLexicon()
const resolver = new GlyphResolver(lexicon) // uses LexicalScorer by default

const matches = await resolver.resolve('buscar un paciente', { limit: 3 })
// → [{ entry: LexiconEntry, score: number }, ...] sorted best-first

LexicalScorer scores token overlap against each glyph's name, tags, and intent (a name/tag hit outweighs an intent hit). No model, no network.

Opt-in — semantic embeddings

pnpm add @huggingface/transformers

import { GlyphResolver, createTransformersScorer } from '@glyphp/resolver'

const scorer = await createTransformersScorer() // downloads MiniLM on first run
const resolver = new GlyphResolver(lexicon, scorer)

EmbeddingScorer ranks by cosine similarity. The embedding function is injected, so any backend works — pass your own EmbedFn to new EmbeddingScorer(embed) to use an API instead.