Windows install

Native Windows installer for Report Anonymizer. Bundles the embedded Python runtime, llama-server.exe (CPU / CUDA / Vulkan variants), pandoc, pdftotext and everything else needed to run the pipeline offline. Tested on Windows 10 (1809+) and Windows 11.

Download

 Report-Anonymizer-Setup-x64-1.0.0.exe  ·  338 MB

The installer is unsigned today (no Authenticode certificate); SmartScreen may flag it on first run with a "Windows protected your PC" dialog. Click More info → Run anyway to proceed. The SHA256 of every release is published in the GitHub Release page so the file can be verified out-of-band.

What the wizard does

The Setup wizard runs entirely without admin rights and writes everything under your user profile:

Path	Content
%LOCALAPPDATA%\Programs\report-anonymizer\	Install root: embedded Python, llama-server.exe, pandoc, pdftotext, GUI launcher
%APPDATA%\report-anonymizer\	Server preset (server.yml), preferences (app_settings.yml), Hugging Face token, installer sentinel
%LOCALAPPDATA%\report-anonymizer\	Downloaded GGUF models, transient cache

1 — Pick the llama.cpp variant

Hardware detection runs nvidia-smi, then PowerShell Get-CimInstance, then legacy wmic. The recommended radio is preselected; you can override by clicking another row.

Three real variants ship in the EXE:

• CPU (AVX2, ~10 MB) — universal fallback. Works on any x86-64 CPU, no GPU needed. Slower for the 4 B-parameter baseline (~30 s on the 5-PDF test) but plenty fast for occasional runs.
• CUDA 12.x (~150 MB) — NVIDIA GPUs with driver ≥ 535 (CUDA 12.0). Best throughput on the 4 B preset (4-5× faster than CPU).
• Vulkan (~30 MB) — NVIDIA / AMD / Intel via the Vulkan runtime. Pick this on AMD / Intel iGPU or when CUDA isn't available.

Two extra radios:

• Skip — install Report Anonymizer without any llama-server. The GUI's Configure deployment page lets you point at an external server later (BYOS mode).
• Keep existing — appears when the installer detects a previous install of one of the three variants. Preserves whatever variant was already on disk; the recommended variant is still pre-selected if it differs from the existing one.

2 — Install to default location

Per-user install. No admin prompt, no privilege escalation.

The default path is %LOCALAPPDATA%\Programs\report-anonymizer\, which doesn't need admin rights. You can override it but the GUI still expects the supporting folders under %APPDATA% and %LOCALAPPDATA%.

3 — First launch

After the wizard finishes, Report Anonymizer opens automatically. Then:

• The first-run wizard detects the bundled llama-server.exe via the installer sentinel (%APPDATA%\report-anonymizer\.installer_choice.json), skips the "install llama-server" step entirely and lets you pick a preset + download the GGUF model in one click.
• The Server panel → Configure deployment dialog now exposes a Backend dropdown (CPU / CUDA / Vulkan) on Windows, so you can switch backend after the install without re-running Setup.
• The GUI lives in %APPDATA%\report-anonymizer\ (config) and %LOCALAPPDATA%\report-anonymizer\ (downloaded models and cache).

Uninstall

Use Settings → Apps → Installed apps → Report Anonymizer → Uninstall. The uninstaller asks one question:

Remove user data too?

• Yes — also delete settings and downloaded models (%APPDATA%\report-anonymizer + %LOCALAPPDATA%\report-anonymizer).
• No (default) — keep them so a future reinstall can reuse the models without re-downloading. The install folder is removed regardless of this choice.

Silent uninstall (/VERYSILENT) defaults to No (keep user data), matching the conservative behaviour of every other Windows uninstaller.

Build from source

The installer is produced by packaging/windows/build.ps1. On a Windows 11 dev box with PowerShell 5.1 / 7.x:

cd packaging\windows
.\build.ps1                       # full build, ~18 min cold
.\build.ps1 -SkipBootstrap        # ~5 min when the tooling cache exists
.\build.ps1 -Clean                # cold rebuild from scratch
.\build.ps1 -Lean                 # CPU-only build, ~225 MB EXE

The script bootstraps 7zr.exe, Inno Setup 6 and MinGW UCRT into packaging\windows\build-cache\tools\, downloads the Python embed ZIP, the WeasyPrint Windows bundle, poppler-windows and the three llama.cpp variant ZIPs, assembles the staging tree under packaging\windows\staging\app\, then runs ISCC to produce packaging\windows\dist\Report-Anonymizer-Setup-x64-<version>.exe.

Use pwsh -ExecutionPolicy Bypass -File .\build.ps1 if PowerShell refuses to run the script under a restricted Execution Policy.