qlaya_flutter

QLaya Flutter is a fast, type-safe Flutter & Dart client for the QLaya on-device decision engine, powered by the qlaya.int8.onnx model hosted on Hugging Face Hub.

It enables intelligent, low-latency decision making directly inside your Flutter and Dart applications:

• 🎯 Intent & Category Classification — Map user inputs to target categories
• ⚡ Continuous Scoring — Evaluate severity, urgency, or sentiment on a 0.0 to 1.0 scale
• ✅ Boolean Verification — Validate hypotheses and statements against input text
• 🔀 Workflow Routing — Automatically route requests and inquiries to appropriate handlers

---

Installation

Add qlaya_flutter to your project dependencies:

flutter pub add qlaya_flutter

Or for pure Dart applications:

dart pub add qlaya_flutter

---

Model Download

The official quantized model (qlaya.int8.onnx, ~571.9 MB) is hosted on Hugging Face Hub at saipy10/qlaya.

You can download it using either the command line or directly inside your Dart/Flutter code:

Option A: Download via Command Line

Run the built-in downloader to save the model to your project:

# Downloads qlaya.int8.onnx to the current directory
dart run qlaya_flutter:download

# Or specify a target directory
dart run qlaya_flutter:download --dir=./models

Option B: Download Programmatically in Flutter

You can verify and download the model weights on-demand with progress updates:

import 'package:qlaya_flutter/qlaya_flutter.dart';

Future<void> initModel() async {
  // Checks local presence and downloads from Hugging Face if missing
  await QLayaDownloader.ensureModel(
    onProgress: (progress) {
      print('Downloading: ${progress.percent}% (${progress.receivedMb} / ${progress.totalMb} MB)');
    },
  );
}

---

Basic Usage

1. Initialize Client

import 'package:qlaya_flutter/qlaya_flutter.dart';

final client = QLayaClient(); // Uses default endpoint or custom baseUrl

2. Intent Classification

Classify input text into one of several predefined choices:

final result = await client.classify(
  text: 'I was charged twice on my monthly invoice, please refund my money',
  choices: ['refund_request', 'subscription_cancel', 'tech_support', 'billing_ops'],
);

print('Choice: ${result.choice}');
print('Confidence: ${result.confidence}');

3. Continuous Scoring / Rating

Rate the urgency, priority, or severity of an incident on a 0.0 to 1.0 scale:

final urgency = await client.score(
  text: 'CRITICAL: Database connection pool completely exhausted in production!',
  instruction: 'Rate the urgency level from 0.0 (low) to 1.0 (emergency)',
);

print('Urgency Score: ${urgency.score}');

4. Boolean Verification

Verify whether a condition or hypothesis holds true for the input:

final verification = await client.verify(
  text: 'The application crashes immediately whenever I tap my profile picture',
  statement: 'User is reporting a software bug',
);

print('Verified: ${verification.value}'); // true

5. Workflow Routing

Route customer requests or messages to the appropriate team or service:

final routing = await client.route(
  text: 'We need to upgrade our plan to 25 enterprise user seats for next quarter',
  routes: ['sales_enterprise', 'tier1_support', 'billing_ops', 'legal_compliance'],
);

print('Target Route: ${routing.route}');

6. Clean Up Resources

client.close();

---

Interactive Flutter Example

This package includes a full Material 3 Flutter application in the example/ directory demonstrating all four decision tasks and model download management:

To run the example app:

cd example
flutter run

---

Model Specifications

Parameter	Specification
Model Variant	qlaya.int8.onnx
Host Repository	Hugging Face (saipy10/qlaya)
Direct Download	https://huggingface.co/saipy10/qlaya/resolve/main/qlaya.int8.onnx
Quantization	INT8 per-channel
Latency (p50)	~134.7 ms
RAM Working Set	~590 MB
Model Size	~571.9 MB

---

License

Apache-2.0 — see LICENSE for details.