PyRagix Win

Native Windows 11 desktop frontend for PyRagix.Net, a local-first RAG (Retrieval Augmented Generation) pipeline. Built with WinUI 3, unpackaged, targeting Windows 11.

Features

• Chat - query your document knowledge base via a local LLM, with a conversational bubble UI
• Ingest - point at a folder and run the full PyRagix.Net ingestion pipeline with real-time progress tracking and cancellation support
• Settings - edit the full PyRagixConfig surface (28 properties across 6 groups), test your LLM connection, and persist changes to settings.toml

Everything runs locally. No cloud APIs, no data leaving your machine.

Prerequisites

1. Windows 11 (build 22621 or later)
2. .NET 10.0 SDK - Download here
3. Visual Studio 2022 or later with the Windows App SDK workload (for the WinUI 3 build toolchain)
4. PyRagix.Net 0.3.2 - available on NuGet.org
5. ONNX models - see ONNX Models below
6. Local LLM server (any OpenAI-compatible endpoint):
   • Ollama
   • LM Studio
   • llamacpp server mode
   • KoboldCpp, vLLM, LocalAI, or any /v1/chat/completions-compatible server

Building

git clone https://github.com/psarno/pyragix-win.git
cd pyragix-win
dotnet restore
dotnet build src/PyRagix.Win/PyRagix.Win.csproj -c Debug

Or open pyragix-win.sln in Visual Studio and build from there. The project targets x64 only (required for FaissNet and ONNX Runtime native binaries).

ONNX Models

PyRagix.Net requires two ONNX models for embedding and reranking. Export them once with Python:

pip install optimum[exporters-onnx]

# Embedding model
optimum-cli export onnx \
  --model sentence-transformers/all-MiniLM-L6-v2 \
  --task feature-extraction \
  src/PyRagix.Win/Models/embeddings

# Reranker model
optimum-cli export onnx \
  --model cross-encoder/ms-marco-MiniLM-L-6-v2 \
  --task text-classification \
  src/PyRagix.Win/Models/reranker

The app copies these to the output directory at build time. On first launch, it writes a settings.toml to %LOCALAPPDATA%\PyRagix\ with model paths pre-set to the output directory locations.

See PyRagix.Net docs/ONNX_SETUP.md for full model export details.

Running

After building, launch the executable directly:

bin\Debug\net10.0-windows10.0.22621.0\win-x64\PyRagix.Win.exe

Or press F5 in Visual Studio.

On first launch, a default settings.toml is written to %LOCALAPPDATA%\PyRagix\. Open the Settings page to configure your LLM endpoint and model name before querying.

Configuration

Settings are stored at %LOCALAPPDATA%\PyRagix\settings.toml. The Settings page in the app provides a full GUI editor. You can also edit the file directly - see src/PyRagix.Win/settings.example.toml for the full schema with comments.

Key settings:

# LLM (any OpenAI-compatible server)
LlmEndpoint = "http://localhost:11434"   # Ollama default
LlmModel    = "qwen2.5:7b"

# Paths (set automatically on first run)
EmbeddingModelPath = "C:/path/to/Models/embeddings/model.onnx"
RerankerModelPath  = "C:/path/to/Models/reranker/model.onnx"

# RAG pipeline features
EnableQueryExpansion   = true
EnableHybridSearch     = true
EnableReranking        = true
EnableSemanticChunking = true

Project Structure

pyragix-win/
├── pyragix-win.sln
└── src/PyRagix.Win/
    ├── App.xaml / App.xaml.cs          # Bootstraps services and ViewModels
    ├── MainWindow.xaml / .cs           # Mica backdrop, DPI-aware sizing
    ├── Program.cs                       # Custom entry point (PerMonitorV2 DPI)
    ├── app.manifest                     # DPI awareness + Windows 11 compatibility
    ├── settings.example.toml           # Configuration template
    ├── Assets/                          # AppIcon.ico, SVG source
    ├── Services/
    │   └── RagService.cs               # Thread-safe RagEngine singleton
    ├── ViewModels/
    │   ├── ChatViewModel.cs
    │   ├── DocumentsViewModel.cs
    │   └── SettingsViewModel.cs
    ├── Views/
    │   ├── ShellPage.xaml / .cs        # NavigationView shell
    │   ├── ChatPage.xaml / .cs         # Chat bubbles + query input
    │   ├── DocumentsPage.xaml / .cs    # Folder picker + ingestion progress
    │   └── SettingsPage.xaml / .cs     # Config editor (SettingsExpander groups)
    ├── Models/                          # ChatMessage.cs, IngestionLogEntry.cs
    │                                    # (embeddings/ and reranker/ dirs are gitignored)
    └── Helpers/
        ├── ConfigSerializer.cs          # Tomlyn write-back for settings
        └── WindowHelper.cs              # HWND + DPI-aware window sizing

Architecture

Standard MVVM using CommunityToolkit.Mvvm v8.4 with partial properties (AOT-compatible, no MVVMTK0045 warnings on WinUI 3).

RagService is a lazy-initialized singleton that owns the RagEngine lifetime. ViewModels call it from background threads via Task.Run and marshal results back to the UI thread with DispatcherQueue.TryEnqueue. IProgress<IngestionProgressUpdate> callbacks follow the same pattern.

The app is unpackaged - no AppX, no MSIX, no Package.appxmanifest. File pickers are initialized with the window HWND via WinRT.Interop.InitializeWithWindow.

Dependencies

Package	Version
Microsoft.WindowsAppSDK	1.8.260317003
CommunityToolkit.Mvvm	8.4.2
CommunityToolkit.WinUI.Controls.SettingsControls	8.2.251219
Tomlyn	0.19.0
PyRagix.Net	0.3.2

License

MIT License - see LICENSE for details.