Use Laya from Node.js

In-process: @receptron/laya

@receptron/laya runs the model with ONNX Runtime, so neither Python nor PyTorch is needed at runtime. It requires Node.js 20 or newer.

npm install @receptron/laya

import { Laya } from "@receptron/laya";

const laya = await Laya.load();

const result = await laya.systemOne(
  { subject: "Refund not received", body: "I cancelled two weeks ago and still have no refund..." },
  {
    department: {
      type: "choice",
      instructions: "Which team should handle this ticket?",
      criteria: { billing: "payments, refunds, invoices", support: "product help and bugs", sales: "new purchases" },
    },
    churn_risk: { type: "noul", instructions: "Is the customer likely to cancel or dispute?" },
  },
);

result.answers.department.choice; // "billing"
await laya.close();

Per its README, the fp32 ONNX weights are about 1.7 GB and download on first use. Budget roughly 2 GB of RAM. A three-question call takes about 140 ms on an Apple Silicon CPU once warm, and the output matches the Python implementation to four decimal places. The package is MIT-licensed; the weights are Apache 2.0.

Over HTTP: laya-studio

laya-studio is a zero-dependency TypeScript client for a running Laya server (Node.js 18.17+):

npm install laya-studio
pip install "laya[serve]"
laya-serve                          # binds 0.0.0.0:8000

Other JavaScript projects

• @r4ai/laya-web: ONNX Runtime Web runtime for browsers and Node.js.
• kevala: a Rust engine compiled to WebAssembly with WebGPU kernels, installable from npm.
• gg-friggin-ez: a profanity and toxicity screener that defaults to Jev and can point at a local Laya endpoint.

Caveats

The English checkpoint's options must fit a 192-token budget and the state is cut at 512 tokens. The receptron README recommends fewer than about 20 options per choice question.