Phi4 AgenticAI WinUI3 Example

Overview

This repository contains a WinUI 3 desktop chat client that runs entirely on-device against a local Phi-4 ONNX model. The UI is implemented with the Windows App SDK and CommunityToolkit.Mvvm, while inference is handled through Microsoft.ML.OnnxRuntimeGenAI with the DirectML execution provider. Microsoft Semantic Kernel maintains a rolling chat history so every response is grounded in previous turns.

Key Components

• ChatAppGenAI/ChatAppGenAI.csproj – WinUI 3 application targeting net8.0-windows10.0.19041.0, referencing Windows App SDK 1.8, CommunityToolkit.Mvvm, Semantic Kernel, and OnnxRuntime GenAI packages.
• ChatAppGenAI/MainWindow.xaml(.cs) – Declarative chat surface plus the VM view-model that wires UI commands to the Phi runner.
• ChatAppGenAI/Phi4Runner.cs – Loads the Phi-4 weights, initializes Semantic Kernel, and exposes GenerateStreamingResponseAsync for token-by-token streaming output.
• ChatAppGenAI/Message.cs – Observable message model that the XAML view binds to for chat rendering.

Prerequisites

• Windows 10/11 build 19041 or later.
• Visual Studio 2022 17.10+ with the ".NET Desktop Development" and "Windows App SDK" workloads, or the .NET 8 SDK plus Windows App SDK runtime.
• DirectX 12–capable GPU/driver for DirectML acceleration.
• Local Phi-4 ONNX model files (see below).

Model Setup

1. Download a Phi-4 ONNX checkpoint that exposes model.onnx, tokenizer assets, and genai_config.json (for example, from the official Microsoft Hugging Face releases).
2. Place the files under ChatAppGenAI/Phi4/ so the layout matches:

   ChatAppGenAI/Phi4/
   ├── model.onnx
   ├── tokenizer.json
   ├── tokenizer_config.json
   ├── special_tokens_map.json
   ├── added_tokens.json
   └── ...
   

3. If you prefer another location, update the pathHead constant in ChatAppGenAI/Phi4Runner.cs to point at your directory.

Build & Run

1. Restore packages: dotnet restore ChatAppGenAI/ChatAppGenAI.csproj.
2. Build for x64 (DirectML requires it): dotnet build ChatAppGenAI/ChatAppGenAI.csproj -c Debug -p:Platform=x64.
3. Launch from Visual Studio (F5) or run ChatAppGenAI/bin/Debug/net8.0-windows10.0.19041.0/win-x64/ChatAppGenAI.exe.
4. When the window opens, wait for the "Model loaded" debug output, type a prompt, and press Enter. Use the Start New Conversation button to clear history and reapply the system message.

Configuration Tips

• Replace the hard-coded pathHead with AppContext.BaseDirectory or expose a settings UI if you need portable builds.
• Phi4Runner.MaxContextTokens governs how aggressively old turns are trimmed; tweak it for smaller GPUs.
• The MVVM layer (VM) centralizes dispatcher access via a helper, which keeps UI updates safe when streaming from background threads.

Troubleshooting

• Model not found – ensure model.onnx exists at Phi4Runner.ModelPath and the process has read permission.
• No GPU acceleration – verify DirectML-compatible drivers are installed; otherwise switch to the CPU provider by changing the NuGet dependency.
• Stalled UI – confirm the dispatcher helper in MainWindow.xaml.cs is receiving a valid DispatcherQueue (already fixed by initializing it before subscribing to events).
• Package restore errors – run nuget locals all -clear, then dotnet restore to refresh the cache.

Customize the runner, prompts, or Semantic Kernel integration to adapt the sample into your own on-device AI assistant.