laya_flutter 0.1.4

Offline multilingual Laya for Flutter: on-device choice, score, and yes/no via ONNX. Open-weight alternative to Jev.

laya_flutter

An offline decision model for Flutter. If you know Jev, Laya is the open-weight model for that same job. Jev is TypeSafe's hosted API. This package runs Laya on the device, and it does not call Jev. Describe a situation in ordinary language, ask a question with a clear shape, and get an answer your app can act on. The first open downloads the published multilingual model. After that, every decision stays on the device.

The picture is the included example: classic Snake. Each step waits for the model to pick a direction. Please open an issue if something breaks or the docs should say more.

How a decision works

You do not ask the model to write a paragraph. You hand it two things:

1. A situation — a short description of what is going on right now.
2. One or more questions — each with an id, a kind of answer, instructions, and the options or scale you want used.

The model returns a typed answer for every question, in one call. Your app reads the choice, the score, or the yes/no and continues. Several questions can share the same situation.

What you can ask

You want	Ask with	You get back
Pick one option you named	LayaQuestionType.choice	The chosen key, such as left
Place the situation on a scale you wrote	LayaQuestionType.score	A level on that scale
A yes or no	LayaQuestionType.noul	How likely “true” is, and which side is stronger

A choice is a door, a direction, a next step in a flow. A score is “how urgent is this?” on levels you defined. A yes/no is a gate: show the screen, or don’t. The wording of the situation and the questions is yours. The published model is multilingual, so that wording can be in the languages it was built for.

What you can build

Anything that needs a small, structured decision and should keep working after the model is on the device:

• A character, opponent, or guide that picks its next action from options you wrote.
• A rating of the current situation on a scale your product already uses.
• A yes/no branch that does not call a server for every tap.
• Several of those at once, about the same moment.

The library does not write essays, and it does not train a new model. It answers the questions you asked.

See it in Snake

The repository includes a playable game in example/. That is the clearest picture of the decision loop:

• Every step, the game describes the board and asks which way to go: up, down, left, or right.
• The snake moves when the model answers. A clock does not push it forward.
• Food appears on an empty cell.
• A wall or the snake’s own body ends the run. A new game starts on its own.

The Snake screen uses a Snake-specific checkpoint that you prepare locally. Your own app can instead call LayaFlutter.open and use the general multilingual model. Setup for the demo is in example/README.md. The bundle path for a device lives in example/lib/app_settings.dart (snakeCheckpointDir).

cd example
flutter run

In Android Studio, open this repository and select the example run configuration ( example/lib/main.dart).

Ask for a decision

dependencies:
  laya_flutter: ^0.1.1

import 'dart:io';

import 'package:laya_flutter/laya_flutter.dart';

Future<void> runLaya(Directory cache) async {
  final LoadedRuntime runtime = await LayaFlutter.open(cache);

  final Map<String, LayaAnswer> answers = await runtime.predict(
    'Player level 3, inventory: key',
    <LayaQuestion>[
      LayaQuestion(
        id: 'door',
        type: LayaQuestionType.choice,
        instructions: 'Which door do you open?',
        criteria: <String, String>{
          'left': 'A wooden door',
          'right': 'An iron gate',
        },
      ),
    ],
  );

  print(answers['door']?.choice);
  await runtime.close();
}

LayaFlutter.open needs the network once, to fetch the multilingual checkpoint. Later opens reuse the cache directory you passed in.

If you already have a prepared checkpoint for a specific task (Snake is one), open that folder instead:

final LoadedRuntime runtime = await LayaFlutter.openLocalBundle(
  Directory('/path/to/bundle'),
);

That folder needs laya.onnx, tokenizer.json, and rl_agent_config.json from the same checkpoint.

Field names and probability maps are documented on LayaQuestion and LayaAnswer.

License

MIT — see LICENSE.