DigUp runs EmbeddingGemma 2 on your Mac to search files by content

A free menu-bar Mac app that indexes your own files with Google DeepMind's EmbeddingGemma 2 appeared on GitHub on 9 October. DigUp is MIT-licensed, weighs about twenty megabytes before the model, and runs everything on the machine it is searching. Its author, Abdur Rahim, says the pitch is simple: describe what you remember and it opens the file at the right moment.

What Happened

DigUp is a native Swift app for Apple Silicon Macs on macOS 14 or newer, wrapping ggml-org's Q8_0 GGUF build of EmbeddingGemma 2 on llama.cpp with Metal. It installs via Homebrew or a signed, notarised DMG. Describe something you half-remember and it opens the file at the right place: a zebra seen in a video plays from that moment, a phrase jumps to that minute of a podcast, a clause brings up the PDF page with your words marked. A query typed in English can find a note written in Bengali or Arabic, and vice versa. With code search enabled, code: retry a failed call with backoff opens the function in your editor.

Per the author's post:

• One-time model download: 865 MB (310 MB text, 555 MB image and audio encoders)
• Loaded while searching: the text encoder only, about 250 MB
• Indexing peak: under 2 GB; the model runs in a process that quits when it finishes
• Latency: results about a tenth of a second after typing stops
• Shortcut: Shift-Cmd-Space from any app

Why This Matters

Privacy is the argument the author makes most plainly. Nothing DigUp indexes leaves the Mac; it goes online once to fetch the model and once a day to check for updates, which sends nothing about your files and can be switched off. There is no account and no telemetry.

The limits deserve stating just as clearly, and most of them come from the author. His own evaluations cover English, Bengali and Arabic because those were the languages of the files he had, and Google notes the model is not equally strong in every language. Matching exact words inside screenshots leans on Apple's text recognition, which knows about 25 languages. Languages written without spaces, such as Chinese, Japanese and Thai, are not split into words. It is Apple Silicon only, and nobody outside the project has measured search quality or the claimed latency. We covered the model release earlier.

What's Next

The repository reports 753 stars, 52 forks and one contributor, and the project is days old: the first commit and the release both landed on 9 October, with contributing guidelines added the day after. Version 0.6.0 is current. What it lacks is an independent benchmark, the kind of number that would turn a promising demo into a claim worth trusting.