@verbx/localize-offline

On-device translation for VerbX Localize — no network, no text leaving the machine. This is the mode for regulated environments: healthcare, banking, government. It is what privacyMode: "offline" has always meant.

npm install @verbx/localize-offline

⚠️ Honest status. The tokenizer is tested against the real vocabulary and the generation loop is tested against scripted inference, but this path has not yet been run against a real model end to end on a device. Talk to us before depending on it.

How it fits

It plugs into VerbXProviderConfig.service — the same seam the Electron SDK uses — so the engine needs no changes: batching, caching, dedupe, the discovery filter and interpolation behave exactly as they do against the cloud.

import { OnnxTranslationService } from "@verbx/localize-offline";

const service = new OnnxTranslationService({
  packs: [{ source: "en", target: "hi", encoder, decoder, tokenizer }],
  tensors: { int64: (data, dims) => /* host's tensor factory */ },
});

// <VerbXProvider config={{ projectKey, privacyMode: "offline", service }}>

What is and isn't in the box

• In: the Unigram tokenizer and the encoder/decoder generation loop — identical on every host.
• Out, deliberately: any onnxruntime import. -node, -react-native and -web are three packages with three binaries; importing one would break the other two hosts. You supply the ONNX runtime sessions and a tensor factory.
• A pack is a downloaded language pair, not bundled into your app — a complete en→hi direction is roughly 111 MiB (int8 OPUS-MT). verifyPack checks a pack's SHA-256 before use, and assertUsable refuses a pack whose licence forbids commercial use.

Docs

Full guide and the offline/on-prem notes: localize.verbxeco.com — the Docs section is readable without an account.

License

Proprietary. See LICENSE.md and the VerbX terms of service.