laya-sdk

Effect-first TypeScript SDK for local Laya decision models (ONNX). Same Decision shapes as a cloud provider, answered on-device — browser or Node. No API keys, no egress.

npm install laya-sdk effect   # effect is a peer dependency (v4)

import { loadLayaRuntime, makeLayaDecisionLive } from "laya-sdk";
import { Decision, DecisionModel } from "effect/unstable/ai";
import { Effect, Schema } from "effect";

const rt = await loadLayaRuntime(); // zip (~434MB) → unzip → cache (IDB or ~/.laya)

const definition = Decision.make({
  input: Schema.String,
  decisions: {
    urgent: Decision.probability({
      instructions: "Is this an urgent production blocker?",
      criteria: { false: "No time pressure", true: "Needs action now" },
    }),
  },
});

const { answers } = await Effect.runPromise(
  DecisionModel.decide(definition, {
    input: "The database disk is 100% full and writes are failing.",
  }).pipe(Effect.provide(makeLayaDecisionLive(rt))),
);
// => { urgent: { probability: 0.93… } }

Options

await loadLayaRuntime({
  modelUrl: "https://my-cdn/laya_int8.onnx", // default: zipped HF int8 release
  tokenizerRepo: "Mattepiu/laya-onnx", // default
  numThreads: 1, // default (no COOP/COEP needed)
  executionProviders: ["wasm"], // default
  onProgress: (label, loaded, total) => {}, // download progress
  cacheDir: "~/.laya", // Node only: default ~/.laya
});

.zip URLs are unzipped transparently (one .onnx entry expected); .gz URLs are gunzipped (see Transfer below). Decoded weights are cached keyed by URL — IndexedDB in browsers, ~/.laya files in Node (LAYA_CACHE_DIR env or cacheDir option overrides) — repeat loads skip the network entirely. clearCachedModel() drops the cached weights (force-refresh a re-released model at the same URL).

Model (measured 2026-09-20)

int8 weights + fp32 embeddings: 434 MB zip transfer, 554 MB decoded. ✅ verified.

Rules: marker dim is fixed at 2 (binary choice and noul only), 512-token budget, English text.

Transfer (no accuracy impact — unzip/gunzip are byte-identical)

Default is laya_int8.zip (434 MB, hosted on HF with CORS + content-length). Self-hosting instead? Either zip the .onnx (one entry) or gzip -9 -k laya_int8.onnx (554 MB -> 434 MB, measured) and serve with CORS

• content-length (progress bar), without content-encoding — the SDK decodes from the file suffix itself. (If your host auto-gzips with content-encoding: gzip, fetch already decoded it and the SDK skips re-decoding.) Brotli is skipped: marginal extra saving, spottier browser support.

Quantization stays on the shelf until an eval harness can prove zero accuracy loss — the transfer win above costs nothing in quality.

Requirements

• effect ^4.0.0-rc.112 (peer), modern browser or Node 22+ / Bun.
• First load downloads + unzips weights into a cache (IndexedDB in browsers, ~/.laya in Node); repeat loads read from disk (no network). Pass a Blob/object URL as modelUrl to bypass download entirely — fetch handles it.

License

MIT