Caches

Everything expensive is persisted under the platform’s LocalApplicationData directory, in DuplicateFileAssistant/: %LOCALAPPDATA%\DuplicateFileAssistant on Windows, ~/.local/share/DuplicateFileAssistant on Linux.

Cache	File	Contents	Invalidated by
File and perceptual hashes	scan-cache.json	SHA-256/MD5 and dHash values	path, length, last write time, algorithm
Embeddings	embeddings-<model>.bin	one float[] per image	path, length, last write time; model id in the file name
Quality scores	quality-<scorer>.bin	one score per image	path, length, last write time; scorer id in the file name
Models	models/<model>/...	ONNX files and .verified SHA-256 markers	SHA-256 of the downloaded file

The scan cache is keyed by path, length, last write time and algorithm. The embedding and quality caches are keyed by path, length and last write time; the model or scorer is part of the file name. A changed length or timestamp recomputes the value.

Not persisted, because they are cheap or derived: file enumeration, the SimHash index, BK-trees, grouping and the final report. They are rebuilt from the caches on every run.

Maintaining the caches

dotnet run --project Cli -- cache --json
dotnet run --project Cli -- cache --prune
dotnet run --project Cli -- cache --clear --apply

Command	Effect
cache	List entries and bytes per cache.
cache --prune	Remove entries whose files no longer exist.
cache --clear --apply	Delete the computed caches. Downloaded models are kept.
cache --cache-dir <dir>	Inspect another cache directory.

duplicates and similar-images accept --prune-cache to prune while scanning. To override individual files, use --cache-path, --embedding-cache and --quality-cache. --no-cache disables the data caches for one run; the model cache is separate and keeps its downloaded files.

The known trade-off

Cache entries are keyed by path, length and last write time. A file modified without changing its size or timestamp (for example with touch -r) can reuse a stale value. Deleting the caches is always safe: the next run recomputes what it needs.

The desktop app uses the same scan cache for its content-hash tabs. The CLI and the app share the cache directory, so a folder scanned by one is faster in the other. In the Similar images tab, the Models and caches expander shows the same totals and offers Refresh, Prune, Clear caches and Open model folder. The model folder itself is chosen in Settings → Models.