DigUp Plus 0.1.5

Windows portable build, 64-bit Windows 10/11. Unpack the zip anywhere and run DigUpPlus.exe — nothing is installed, and settings, logs, the index and the model live in the data folder beside it. The build isn't signed: choose More info → Run anyway if Windows asks. The zip is ~470 MB, mostly the CUDA runtime the helper carries so the model runs on NVIDIA cards.

SHA-256 of DigUpPlus-0.1.5-windows-x64.zip: 8e74d320f4a341c5ca1a9c694cbd4d49687cdf4248d9c85bb3690a0cf3d9acc

What is in 0.1.5

The window behaves the way it reads. The X quits DigUp Plus outright — no forgotten half-running instances; indexing pauses and resumes on the next start. Esc (after everything else there is left to cancel) and Ctrl+W put the window away into the tray instead, with the app and its indexing going on.

The views tell the truth. The window opens on Images — the whole picture ranking in cosine layers, paged as deep as it goes. All files (the mixed-score view) is the last chip: its numbers come from a different ranking, and as the default it used to hide hundreds of cosine matches behind the gate. The Media chip went — its kinds have their own chips next to Images. When the open filter has nothing but other views do, the empty state points at them with counts.

Find similar (search by example). Right-click a picture, screenshot, video or audio result → Find similar: its own embedding becomes the query, the ranking is by cosine, and the reference itself stays out of its results. No re-embedding, no cloud. A failed lookup is a visible error, never a keyword search in disguise.

Small corpora search right. A meaning match on a tiny index could never pass the relevance gate (three files measured z 1.41 against the 2.0 threshold — a mathematical ceiling). Under 8 scale-setting vectors a clearly matching file is admitted by an absolute cosine floor, while unrelated queries stay rejected.

File operations at scale (a second review pass, verified by the full test suites): a 5,000-file move now travels in chunks of 500 index updates instead of 5,000 round trips; the destination folder is read once per batch instead of per file; a copy that fails before creating its target no longer reports a "partial copy" that doesn't exist; selecting results no longer repaints every cell on every key.

Datasets as a setting. Folders that look like datasets (thousands of similar files) are still skipped by default — Settings → Folders can index them after all, for collections whose subfolders trip the rule.

The first run opens Settings, not a wizard; nothing downloads without asking. No telemetry, no cloud, no update checks — the only network call is the one-time model download.

A Windows-focused fork of ARahim3/DigUp; see the README for what it changes.