FAQ
Does recognition need a network connection?
No. Once models are installed, recognition runs entirely on-device. Networking is only used for catalog refresh, model installation and update checks. Offline, cached catalogs and installed models keep working.
Are my documents uploaded?
No. The app has no upload path for documents; results stay in the local data directory. Update checks send only the version and app identifier.
Where are models stored, and can I change the disk?
Models, runtimes, caches and logs live in the app data directory:
| Platform | Path |
|---|---|
| Windows | %AppData%\Glyphix |
| macOS | ~/Library/Application Support/Glyphix |
| Linux | $XDG_CONFIG_HOME/Glyphix |
Settings → Storage can move the directory; installed models are moved rather than downloaded again.
Downloads keep dropping — what now?
Check the proxy first: the app uses Settings → Network, and changes apply to new downloads immediately. Transfers resume after a restart or mirror switch, and verified files are reused. When an address keeps failing, the app falls back through the catalog's mirrors and logs each failure.
How many times do I install the shared runtime?
Once. MiniCPM-OCR, Xiaomi-OCR-0 and PaddleOCR-VL share one llama.cpp runtime. Removing it affects every referencing engine, and the confirmation dialog lists them.
Is arm64 supported?
Yes. Windows arm64, Apple Silicon macOS and Linux arm64 all have builds and model packs. Some second-tier capabilities (tray, system decorations) may degrade, and every degraded item is labelled in the engine center and settings.
Can I import PDFs?
Yes, PDFs are imported and recognised page by page. When a page has no renderable image, the image view shows boxes with a notice while other views remain usable.
Why does exported Markdown differ from the preview?
Normally they share one source. Differences usually mean an edit was never committed (commits happen on blur) or the crop cache was cleared. Commit the edit and export again.
Does changing the UI language require a restart?
No. Strings refresh immediately and running batches are unaffected. Recognition language is a separate setting controlled by the batch options.
Can several computers share one model directory?
You can point storage at an external or network drive, but runtimes and model packs are distributed per RID — a Windows pack cannot serve macOS. Install per machine, or prepare one directory per platform.