Nerve (JS/TS) — on-device AI in WebAssembly

A real Transformer text generator and a MiniLM sentence-embedding model, running entirely in WebAssembly — in the browser or Node. No server, no GPU, no API key, no per-token cost. The data never leaves the machine.

Built from Nerve: zero-dependency neural networks in pure C, compiled to a ~65 KB WASM engine.

Install

npm install @fkkarakurt/nerve

Use

import Nerve from "@fkkarakurt/nerve";

const nerve = await Nerve.load();

// 1. generate text (streams)
const story = nerve.generate("Once upon a time", {
  steps: 120,
  onToken: t => process.stdout.write(t),
});

// 2. understand meaning
nerve.similarity("a puppy on the grass", "a young dog in the park"); // ~0.7

// 3. learn your own categories (on-device, in ms)
nerve.teach([
  { text: "schedule a meeting",     label: "calendar" },
  { text: "i want a hamburger",     label: "food" },
  { text: "go for a run",           label: "fitness" },
]);
nerve.classify("i'm hungry for pizza"); // { label: "food", confidence: 0.8, scores: {...} }

// 4. semantic search over your own notes
nerve.index(["The capital of France is Paris.", "Coffee contains caffeine."]);
nerve.search("what keeps me awake?"); // [{ text: "Coffee contains caffeine.", score: 0.4 }, ...]

API

| Method | Description | |--------|-------------| | Nerve.load() | Load the models; returns a ready instance. | | embed(text) | Sentence → L2-normalised Float32Array. | | similarity(a, b) | Cosine similarity (-1..1). | | generate(prompt, opts) | Generate text; opts.onToken streams. | | teach(examples) | Train a classifier on {text, label} examples. | | classify(text) | { label, confidence, scores }. | | index(notes) / search(query, k) | Semantic search over your notes. |

Generation is synchronous (runs to completion); for long outputs in the browser, call it from a Web Worker to keep the UI responsive.

License

GPL-3.0-or-later — see the main repository.