/* Next Match compute core: port of match_core.py (tokenizer, text/image encoders, fusion, ranking). */ (function (root, factory) { if (typeof module === "object" && module.exports) module.exports = factory(); else root.Core = factory(); })(typeof self !== "undefined" ? self : this, function () { "use strict"; const ASSETS = ["model.json", "catalog.json"]; function lowerTr(s) { return s.replace(/I/g, "ı").replace(/İ/g, "i").toLowerCase(); } function tokenize(text, vocab) { const clean = Array.from(lowerTr(text).replace(/-/g, "")).map(ch => (/[\p{L}\p{N}]/u.test(ch) || ch === " " ? ch : " ")).join(""); const ids = []; for (const w of clean.split(/\s+/).filter(Boolean)) { if (w in vocab) { ids.push(vocab[w]); continue; } const chars = Array.from(w); for (let k = chars.length - 1; k > 3; k--) { const p = chars.slice(0, k).join(""); if (p in vocab) { ids.push(vocab[p]); break; } } } return ids; } function textVector(ids, emb) { const [V, D] = emb.shape, out = new Float64Array(D); if (!ids.length) return out; for (const i of ids) for (let d = 0; d < D; d++) out[d] += emb.data[i * D + d]; for (let d = 0; d < D; d++) out[d] /= ids.length; return out; } function l2n(x) { let s = 0; for (const v of x) s += v * v; const n = Math.max(Math.sqrt(s), 1e-12); return Float64Array.from(x, v => v / n); } function init(assets) { const m = NLRT.decode(assets["model.json"]), cat = assets["catalog.json"]; const keys = m.tensors.gallery_keys; return { m, vocab: m.meta.vocab, cat, keys, D: keys.shape[1] }; } function imageFeatures(ctx, pixels, net) { // pixels: Float64Array 3*48*48 in [0,1] (CHW) return NLRT.forward(ctx.m.nets[net], NLRT.T([3, 48, 48], pixels)).data; } function encodeQuery(ctx, pixels, text, mode) { const n = ctx.m.nets, t = ctx.m.tensors; if (mode === "combined") { const fi = imageFeatures(ctx, pixels, "img"); const ft = NLRT.forward(n.txt, NLRT.T([t.emb.shape[1]], textVector(tokenize(text, ctx.vocab), t.emb))).data; const cat = new Float64Array(fi.length + ft.length); cat.set(fi); cat.set(ft, fi.length); const delta = NLRT.forward(n.fuse, NLRT.T([cat.length], cat)).data; return l2n(fi.map((v, i) => v + delta[i])); } if (mode === "image") { const fi = imageFeatures(ctx, pixels, "img_only"); return l2n(NLRT.forward(n.img_only_head, NLRT.T([fi.length], fi)).data); } return l2n(NLRT.forward(n.txt_only, NLRT.T([t.emb_only.shape[1]], textVector(tokenize(text, ctx.vocab), t.emb_only))).data); } function rank(ctx, q, mode) { const K = mode === "image" ? ctx.m.tensors.gallery_keys_img_only : mode === "text" ? ctx.m.tensors.gallery_keys_txt_only : ctx.keys; const N = K.shape[0], D = K.shape[1], s = new Float64Array(N); for (let i = 0; i < N; i++) { let a = 0; for (let d = 0; d < D; d++) a += K.data[i * D + d] * q[d]; s[i] = a; } return Array.from({ length: N }, (_, i) => i).sort((a, b) => (s[b] - s[a]) || (a - b)).map(i => [i, s[i]]); } function runCase(ctx, input) { const px = Float64Array.from(input.pixels, v => v / 255); const out = {}; for (const mode of ["combined", "image", "text"]) { const q = encodeQuery(ctx, px, input.text, mode); out[mode] = rank(ctx, q, mode).slice(0, 5).map(r => r[0]); out[mode + "_q0"] = q[0]; } out.tokens = tokenize(input.text, ctx.vocab); return out; } return { ASSETS, init, tokenize, encodeQuery, rank, runCase }; });