src/embed.js

// ----------------------------------------------------------------------------
// Sentence embeddings with all-MiniLM-L6-v2 running locally in Node via
// Transformers.js (@xenova/transformers). No API key, no external calls — the
// model (~90MB) is downloaded once on first use and cached on disk.
//
// Output: 384-dim, mean-pooled, L2-normalised vectors (cosine == dot product).
// ----------------------------------------------------------------------------

import { pipeline, env } from '@xenova/transformers';

// Cache models under ./.models so repeated deploys don't re-download.
env.cacheDir = process.env.MODEL_CACHE_DIR || './.models';
// We only need inference; disable local-file-only so it can fetch once.
env.allowRemoteModels = true;

let extractorPromise = null;

function getExtractor() {
  if (!extractorPromise) {
    extractorPromise = pipeline('feature-extraction', 'Xenova/all-MiniLM-L6-v2');
  }
  return extractorPromise;
}

/** Warm the model at boot so the first request isn't slow. */
export async function warmup() {
  const e = await getExtractor();
  await e('warmup', { pooling: 'mean', normalize: true });
}

/** Embed a single string -> number[] (384). */
export async function embed(text) {
  const e = await getExtractor();
  const out = await e(String(text ?? ''), { pooling: 'mean', normalize: true });
  return Array.from(out.data);
}

/** Embed many strings -> number[][]. */
export async function embedBatch(texts) {
  const e = await getExtractor();
  const vecs = [];
  for (const t of texts) {
    const out = await e(String(t ?? ''), { pooling: 'mean', normalize: true });
    vecs.push(Array.from(out.data));
  }
  return vecs;
}