The moment, the minute, the page
A video opens at the moment the zebra shows up, a podcast at the minute they talk about sleep, and the preview shows the PDF page that answers. Code opens in your editor at the line.
Describe it.
Dig it up.
Search your computer by what's in your files. DigUp runs Google DeepMind's EmbeddingGemma 2 on your own machine, so a few words find the photo, the PDF page, the minute in a recording or the moment in a video.
Version 0.6.2 · Apple Silicon, macOS 14 or later Version 0.7.0-beta.3 · 64-bit Windows 10 or 11 · in beta Version 0.7.0-beta.3 · .deb or AppImage, x86-64 · in beta Free and open source
zebra in a video
Video
Masai Mara - day 2.mp4 0:42~/Movies/Kenya 2025Images
IMG_2214.jpg~/Pictures/Kenya 2025
IMG_2219.jpg~/Pictures/Kenya 2025
Masai Mara - day 2.mp4
~/Movies/Kenya 2025
Matches the frame at 0:42
3 results↩ Open⌘↩ RevealCtrl+↩ Show␣ Quick Look at 0:42
Audio
The Long Run - Ep 14.mp3 7:55~/Music/Podcasts7:55 – 8:25 of 13:43
The Long Run - Ep 14.mp3
~/Music/Podcasts
Matches what's heard at 7:55
1 result↩ Open⌘↩ RevealCtrl+↩ Show␣ Quick Look at 7:55
Documents
Lease - 14 Elm Street.pdf p. 4HARBOR LANE PROPERTIES RESIDENTIAL LEASE …
Lease - 14 Elm Street.pdf
~/Documents/Home
Matches the text on page 4
HARBOR LANE PROPERTIES RESIDENTIAL LEASE AGREEMENT 12. PETS …
1 result↩ Open⌘↩ RevealCtrl+↩ Show␣ Quick Look at p. 4
Images
IMG_4133.jpg~/Pictures/2025 Summer
IMG_4128.jpg~/Pictures/2025 Summer
IMG_4127.jpg~/Pictures/2025 SummerVideo
IMG_4141.MOV 0:12~/MoviesAudio
Dog barking.m4a~/Music/Field recordings
IMG_4133.jpg
~/Pictures/2025 Summer
Matches the picture
6 results · 2 weaker↩ Open⌘↩ RevealCtrl+↩ Show␣ Quick Look
Video
IMG_4141.MOV 0:12~/MoviesImages
IMG_4127.jpg~/Pictures/2025 SummerIMG_4133.jpg~/Pictures/2025 SummerAudio
Dog barking.m4a~/Music/Field recordings
IMG_4141.MOV
~/Movies
Matches the frame at 0:12
6 results↩ Open⌘↩ RevealCtrl+↩ Show␣ Quick Look at 0:12
Code
retry.ts L1–47// Retry a failed call, waiting longer each time: e…
retry.py L1–33…random import time def retry_with_backoff(a…
retry.test.ts L1–22weather-app/test
test_retry.pybilling-service/tests
5 weaker matches
retry.ts
weather-app/src/net
Matches lines 1–47
4 results · 5 weaker↩ Open⌘↩ RevealCtrl+↩ Show␣ Preview
Opens the clip at the moment the zebra shows up.
Opens the clip at the moment the zebra shows up.Jumps to the minute of the podcast where they talk about sleep.Shows the page of the lease with the clause, among eight pages.Finds the photo by what's in it, and the video and the recording that fit.The same beach, found in Bengali, in Arabic and in English.Finds the function by what it does, and opens it in your editor at the line.⇧⌘Space CtrlShiftSpace opens it from any app
Prefer to watch? DigUp in one minute, with narration.
1:00
One model, one space for every kind of file
EmbeddingGemma 2 is the embedding model Google DeepMind released in October 2026. It turns text, pictures, sound and video into points in one shared space, where things that mean the same land close together, in more than 100 languages. It's open (Apache 2.0) and small enough, at 740M parameters, to run on a laptop.
the beachvideo frames + sound + photossafarivideo frames + photosa podcastsoundhome and foodphotospages and notestext + screenshots
a photo0.74
IMG_4133.jpg
a photo0.73
IMG_4128.jpg
0:25
a sound0.73
Dog barking.m4a
a dog barking on the beach
picturesvideo framessoundtext164 real EmbeddingGemma 2 vectors of the demo's files (768-d, shown in 2-D)
- a photo IMG_4133.jpg 0.74
- a photo IMG_4128.jpg 0.73
- a sound Dog barking.m4a 0.73
It takes you to the spot
Every page, passage, video frame and stretch of sound has its own place in the index, so a result opens where it matched.
Exact words, too
Invoice numbers, error codes and names are matched as written, next to the meaning. On a Mac that includes the text inside screenshots and scans.
In any of more than 100 languages
A search in one language finds files written in another. The evals cover English, Bengali and Arabic; the model knows many more.
Code, if you want it
Turn it on for your projects, then start a search with code: and say what the code does, or type a name to go to where it's defined. Code has an index of its own and never mixes with your files.
Long files, read to the end
Recordings and videos of any length are read whole, 30 seconds of sound and one frame per shot at a time. Big PDFs and documents get their first pages right away and the rest later.
Every kind of file you keep
Pictures, PDFs (scans too), Word, Markdown and other documents, spreadsheets and slides, recordings and videos. The full list
Keeps up quietly
After the first pass, DigUp reads only what's new or changed, a moment after it lands. The model runs in a helper that quits when it's done.
Made for the keyboard
Press the shortcut in any app, type, and you're there. Everything else is a key away.
- ⇧⌘Space CtrlShiftSpaceOpens the search panel from any app
- ↩ EnterOpens the result at its page, moment or line
- Space SpaceA big preview, right there
- ⌘↩ CtrlEnterShows it in FinderShows it in ExplorerShows it in your file manager
- ⌘O CtrlOThe same search in a window, with filters
Nothing you index leaves your computer
The model, the index and every search stay on your machine. DigUp goes online for two things only, and neither carries anything about your files. No account, no telemetry.
Your computer
Your folders
DigUpwith EmbeddingGemma 2
One index file
DigUpwith EmbeddingGemma 2
One index file
OnceThe model, 865 MB, from Hugging Face
DailyAn update check on GitHub, which you can turn off
On a Mac, on Windows and on Linux
The search engine is the same code on every system, with the same model and the same results. The app around it belongs to each.
Mac
Stable0.6.2 · October 21, 2026
Needs an Apple Silicon Mac (M1 or later) with macOS 14 or later.
- In the menu bar. Runs the model on the GPU with Metal.
- Screenshots and scanned pages are also searchable by their text, with Apple's text recognition.
- Videos open at the moment in QuickTime Player or IINA.
Windows
Beta0.7.0-beta.3 · October 22, 2026
Needs 64-bit Windows 10 or 11.
- In the tray. Runs the model on the graphics card with Vulkan when that's faster, otherwise on the processor.
- Installs for you only, with no admin rights.
- Not yet: text inside screenshots, old .doc files, RAW photos, SVG pictures.
Linux
Beta0.7.0-beta.3 · October 22, 2026
Needs a 64-bit Intel or AMD processor: Ubuntu 22.04, Debian 12, Mint 21 or newer for the .deb, and most distributions for the AppImage.
- In the tray. Runs the model on the graphics card with Vulkan when that's faster, otherwise on the processor.
- A .deb for Ubuntu, Debian and Mint, and an AppImage for the rest.
- Not yet: text inside screenshots, old .doc files, RAW photos, SVG pictures.
Watch the same searches on each systemWhat differs between themAll releases on GitHubBuild from source
Questions people ask
More in the FAQ, and fixes for common problems in Troubleshooting.
What is DigUp?
A search app for the files on your computer. It runs Google DeepMind's EmbeddingGemma 2, a model that puts text, pictures, sound and video in one shared space, on your own machine. You describe what you remember, and it finds the photo, the PDF page, the minute in a recording or the moment in a video. It's free and open source, on a Mac, on Windows and on Linux.
How is it different from Spotlight or Windows Search?
They find files by their names, text and metadata. DigUp also finds them by what they show or say, and takes you to the page or moment that matched. They work fine side by side.
What does DigUp send over the network?
Nothing you index leaves your computer. DigUp goes online for two things: to download the model once, from Hugging Face, and to check for updates once a day, on GitHub. The update check sends nothing about your files, and you can turn it off in Settings → General. No account, no telemetry. Privacy and security has the details.
Why is the download 865 MB?
That's the model: the text part (310 MB) and the image and audio encoders (555 MB). It downloads once. The app itself is about 20 MB on a Mac, and 40 to 140 MB on Windows and Linux, depending on the package, since those carry their own PDF, video and image readers.
Will it slow my computer down?
Only the first pass is heavy, and it happens once. After that DigUp reads only what's new or changed. The model runs in a separate helper that quits when it's done, and the search side only loads while you search. On a Mac, Low Power Mode and a hot Mac make the first pass wait, and Index only at set hours can keep it to the night.
Which languages does it understand?
The model knows more than 100 languages, and DigUp searches in all of them, across languages too: a search in English can find a note in Bengali or Arabic. DigUp's tests cover English, Bengali and Arabic because those are the languages of the files it was tested on, and Google notes that the model isn't equally strong in every language. Two limits apply to exact words, not to meaning: the text inside screenshots is read only on a Mac, by Apple's text recognition (about 25 languages), and languages written without spaces, like Chinese, Japanese and Thai, aren't split into words.