shim.js

// JS runtime for the exported moodring model. Mirrors the sklearn pipeline:
// tokenize (runs of >=2 word chars, lowercased), unigrams+bigrams, sublinear
// tf, idf, l2 norm, then per-label logistic scores. Works in browser and Node.

const TOKEN_RE = /[\p{L}\p{N}_]{2,}/gu;

function parseModel(buffer) {
  const bytes = new Uint8Array(buffer);
  const magic = new TextDecoder().decode(bytes.subarray(0, 9));
  if (magic !== "MOODRING1") throw new Error("bad model file");
  const jsonLen = new DataView(buffer).getUint32(9, true);
  const meta = JSON.parse(new TextDecoder().decode(bytes.subarray(13, 13 + jsonLen)));
  const n = meta.n_features;
  let off = 13 + jsonLen;
  const idf = new Float32Array(buffer.slice(off, off + 4 * n));
  off += 4 * n;
  const coef = new Float32Array(buffer.slice(off, off + 4 * n * meta.emotions.length));
  const vocab = new Map(meta.terms.map((t, i) => [t, i]));
  return { meta, idf, coef, vocab, n };
}

function tokenize(text) {
  return (text.toLowerCase().match(TOKEN_RE) || []);
}

function vectorize(model, text) {
  const tokens = tokenize(text);
  const grams = tokens.slice();
  for (let i = 0; i + 1 < tokens.length; i++) grams.push(tokens[i] + " " + tokens[i + 1]);
  const counts = new Map();
  for (const g of grams) {
    const idx = model.vocab.get(g);
    if (idx !== undefined) counts.set(idx, (counts.get(idx) || 0) + 1);
  }
  let normSq = 0;
  const entries = [];
  for (const [idx, tf] of counts) {
    const w = (1 + Math.log(tf)) * model.idf[idx];
    entries.push([idx, w]);
    normSq += w * w;
  }
  const norm = Math.sqrt(normSq) || 1;
  return entries.map(([idx, w]) => [idx, w / norm]);
}

function scores(model, text) {
  const x = vectorize(model, text);
  const { emotions, intercept } = model.meta;
  const out = {};
  for (let j = 0; j < emotions.length; j++) {
    let margin = intercept[j];
    const row = j * model.n;
    for (const [idx, w] of x) margin += model.coef[row + idx] * w;
    out[emotions[j]] = 1 / (1 + Math.exp(-margin));
  }
  return out;
}

function predict(model, text) {
  const s = scores(model, text);
  const { emotions, thresholds } = model.meta;
  let labels = emotions.filter((e, j) => s[e] >= thresholds[j]);
  const fallback = labels.length === 0;
  if (fallback) {
    labels = [emotions.reduce((a, b) => (s[a] >= s[b] ? a : b))];
  }
  const rounded = {};
  for (const e of emotions) rounded[e] = Math.round(s[e] * 10000) / 10000;
  const thr = {};
  emotions.forEach((e, j) => (thr[e] = Math.round(thresholds[j] * 100) / 100));
  return { labels, fallback, scores: rounded, thresholds: thr };
}

// ---- static-Space shim: answer the app's own API routes client-side ----
(function () {
  const realFetch = window.fetch.bind(window);
  let modelPromise = null, testPromise = null;
  const getModel = () => (modelPromise ||= realFetch("./model.bin")
    .then((r) => { if (!r.ok) throw new Error("model download failed"); return r.arrayBuffer(); })
    .then(parseModel));
  const getTest = () => (testPromise ||= realFetch("https://raw.githubusercontent.com/google-research/google-research/master/goemotions/data/test.tsv")
    .then((r) => { if (!r.ok) throw new Error("sample fetch failed"); return r.text(); })
    .then((t) => t.trim().split("\n").map((line) => line.split("\t"))));
  window.fetch = async function (url, opts) {
    const u = typeof url === "string" ? url : url.url;
    if (u.startsWith("/predict")) {
      const model = await getModel();
      const { text } = JSON.parse(opts.body);
      return Response.json(predict(model, text));
    }
    if (u.startsWith("/sample")) {
      const [rows, model] = await Promise.all([getTest(), getModel()]);
      const raw = parseInt(new URLSearchParams(u.split("?")[1]).get("index") || "0", 10);
      const index = Math.max(0, Math.min(rows.length - 1, raw));
      const [text, labels] = rows[index];
      const gold = labels.split(",").map((s) => model.meta.emotions[parseInt(s, 10)]);
      return Response.json({ index, total: rows.length, text, gold });
    }
    return realFetch(url, opts);
  };
})();