@johnhenry/aimatey-native-laya

Run Aimatey typed-decision requests against ConvAI's Laya model on-device, via @receptron/laya -- a real TypeScript/ ONNX Runtime port, not a hosted-API wrapper. No network call, no API key.

Part of the aimatey monorepo.

Requirements

• Node.js 18+ (this package: Node 26+)
• The optional native binding: npm install @receptron/laya
• ~1.7GB of ONNX weights, downloaded from HuggingFace on first use and cached locally (cacheDir, or @receptron/laya's own LAYA_CACHE default)

Installation

npm install @johnhenry/aimatey-native-laya @receptron/laya

Quick Start

import { Bridge } from '@johnhenry/aimatey-core';
import { LayaFrontendAdapter } from '@johnhenry/aimatey-frontend';
import { LayaBackendAdapter } from '@johnhenry/aimatey-native-laya';

const backend = new LayaBackendAdapter();

// Sessions are loaded lazily on first decide(); initialize eagerly to
// surface a missing-dependency or download failure early.
await backend.initialize();

const bridge = new Bridge(new LayaFrontendAdapter(), backend);

// IR in, IR out. The frontend is not involved.
const response = await bridge.decide(
  { headline: 'Local sea levels rising faster than predicted' },
  { urgency: { type: 'score', instructions: 'How urgent is this?', criteria: ['low', 'medium', 'high'] } }
);

// Laya's own request/response shapes in and out, via the frontend adapter.
const native = await bridge.decideFrom({
  state: { headline: 'Local sea levels rising faster than predicted' },
  questions: { urgency: { type: 'score', instructions: 'How urgent is this?', criteria: ['low', 'medium', 'high'] } },
});

await backend.close();

What the adapter declares

• decisionModels: english, multilingual, typed-decisions
• decisionTypes: choice, score, noul
• decisionLimits: 20 choice options (Laya is weak beyond that), 10 score levels, 512 state tokens (the English checkpoint; multilingual allows 1024), no images (decisionImages: false; images are dropped with a capability-unsupported warning)

Behaviour notes

• decide(request, signal) honours signal before loading, before running, and once the result arrives. An in-flight ONNX run cannot be cancelled.
• The checkpoint is chosen when the session loads (subfolder), and @receptron/laya's systemOne() takes no task/lang. A parameters.model that differs from subfolder, and parameters.custom.task/lang, are reported as parameter-unsupported warnings on response.metadata.warnings rather than ignored silently.
• Laya's per-answer rl_agent sub-object is kept under response.raw.rl_agent, keyed by question. Convai measured its act_probability as uninformative, so treat it as diagnostic.
• Laya reports no confidence for noul answers; the adapter derives max(p, 1 - p).

License

MIT - see LICENSE for details.