Keyword search has a fundamental mismatch with how people actually remember things.
You don't remember filenames. You remember content — "that recording where someone mentioned data residency." The filename won't tell you it's at the 37-minute mark.
DigUp (2125 stars, MIT, Swift) addresses exactly that gap: it indexes your Mac by what's inside the files, not by what they're called.
How it works
It uses Google DeepMind's EmbeddingGemma 2, a multimodal embedding model, to map images, PDFs, documents, audio, and video into a single semantic space. You describe what you're looking for in plain language, and it resolves to:
| File type | Granularity |
|---|---|
| Video / audio | Jumps to the timestamp |
| Locates the page containing the match | |
| Code | Jumps to the specific line |
| Images / docs | Semantic content match |
Indexing runs entirely on-device. Files are never uploaded.
Install
brew install --cask arahim3/tap/digup
First launch: downloads the model (865 MB, one time), you pick folders to index, wait for the initial pass, then hit ⇧⌘Space anywhere to query.
Requirements and real limits
- macOS 14+ on Apple Silicon only — Intel Macs are not supported
- The 865 MB model download may need acceleration on some networks (no mirror documented)
- Initial indexing is slow: minutes for hundreds of files, up to an hour for large directories
- It's a personal tool — no permissions, multi-user, or audit. Not an enterprise knowledge base
Why I find it worth writing about
It makes the retrieval layer of RAG tangible.
Enterprise teams spend real money on internal knowledge bases that still return poor results — usually because the system is still matching keywords, not meaning. Ask for "how customer complaints are handled" and you get a pile of directory pages and policy notices containing the words, not the runbook that actually explains the process.
DigUp lets anyone experience the difference between guessing the word and describing the meaning — on their own files, in a few minutes. As a demo for non-technical stakeholders, it does more than a slide deck.
I localized the README to Chinese with a quick-start section covering install, hardware requirements, model download, and indexing time expectations: https://github.com/yangshun2005/DigUp-cn
If it's useful, a star on the original repo supports the author.
Top comments (0)