qlaya_flutter 0.5.0

QLaya Flutter & Dart client for the fast on-device decision engine powered by the qlaya.int8.onnx model for classification, scoring, verification, and routing tasks.

qlaya_flutter

QLaya Flutter — Fast, local, on-device decision engine client for Flutter and Dart, powered by the qlaya.int8.onnx quantized model hosted on Hugging Face Hub.

Hugging Face Model | GitHub Repository | pub.dev Package | Issues & Feedback

---

Model Specifications

The official quantized model weights are hosted on Hugging Face Hub at saipy10/qlaya:

• Model File: qlaya.int8.onnx (~571.9 MB)
• Quantization: ONNX INT8 per-channel
• Latency (p50): ~134.7 ms
• RAM Working Set: ~590 MB
• Direct Download URL: https://huggingface.co/saipy10/qlaya/resolve/main/qlaya.int8.onnx
• Capabilities: Fast on-device classification, continuous scoring, boolean verification, and workflow routing.

---

Installation

Add to your Flutter or Dart project:

flutter pub add qlaya_flutter
# or for Dart CLI / server apps:
dart pub add qlaya_flutter

---

Downloading the Model

You can obtain the model weights either via a command or directly within your code:

1. Download via CLI Command

Run the built-in downloader to fetch qlaya.int8.onnx from Hugging Face into your current directory or custom path:

# Download into current directory
dart run qlaya_flutter:download

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

# Force re-download
dart run qlaya_flutter:download --force

2. Download / Ensure Programmatically in Dart / Flutter

You can also ensure the model exists or download it on-the-fly directly inside your app:

import 'package:qlaya_flutter/qlaya_flutter.dart';

void main() async {
  // Check if model is already downloaded
  if (!QLayaDownloader.isModelAvailable()) {
    print('Downloading model weights from Hugging Face...');
    await QLayaDownloader.download(
      onProgress: (p) {
        print('${p.percent}% (${p.receivedMb} / ${p.totalMb} MB)');
      },
    );
  }
}

---

Inference Usage

Use QLayaClient to run type-safe decision tasks:

import 'package:qlaya_flutter/qlaya_flutter.dart';

void main() async {
  // Connects to local server (defaults to http://localhost:8000)
  final client = QLayaClient();

  // Task 1: Intent Classification
  final intent = await client.classify(
    text: 'I was charged twice, please refund',
    choices: ['refund_request', 'subscription_cancel', 'tech_support'],
  );
  print('Classified Choice: ${intent.choice}');

  // Task 2: Continuous Scoring / Rating
  final urgency = await client.score(
    text: 'CRITICAL: Database connection timeout in production!',
    instruction: 'Rate urgency from 0.0 (low) to 1.0 (emergency)',
  );
  print('Urgency Score: ${urgency.score}');

  // Task 3: Boolean Verification
  final isComplaint = await client.verify(
    text: 'Product stopped working after 2 days.',
    statement: 'Customer is filing a complaint',
  );
  print('Is Complaint: ${isComplaint.value}');

  // Task 4: Workflow Routing
  final routing = await client.route(
    text: 'Need to add 15 enterprise seats for next quarter',
    routes: ['sales_enterprise', 'tier1_support', 'billing_ops'],
  );
  print('Route To: ${routing.route}');

  client.close();
}

---

Running the Inference Server

Start the local QLaya engine server using the downloaded ONNX model:

pip install qlaya[serve]
qlaya-serve --model qlaya.int8.onnx --port 8000

---

License

Apache-2.0 — see LICENSE.