Inside DigUp: How a 740M-Param Multimodal Embedding Model Enables Local Semantic File Search on Mac
xiaohu · x · 2026-10-14
DigUp, open-sourced by developer Abdur Rahim, is a Mac menubar search tool built on Google DeepMind's EmbeddingGemma 2 (Apache 2.0). It searches photos, documents, code, and audio/video clips in chosen folders, jumping straight to the exact video moment, PDF page, or code segment — all processed locally on Apple Silicon with no file uploads.
How one query finds photos and recordings: The model totals 740M params — a 270M text backbone, 170M vision encoder, and 300M audio encoder — mapping text, images, video frames, and audio slices into a shared 768-dim vector space, where text queries rank candidates by cosine similarity.
Why not OCR + speech-to-text: Text-mediated pipelines miss non-textual cues like tone, background sounds, and composition; EmbeddingGemma 2 uses native multimodality instead. It also employs task instruction prefixes (e.g. task: search result | query: ..., task: code retrieval | ...) to align vectors per retrieval scenario, and its decoupled encoders mean pure text search only loads the text backbone.
More from Apps
- Gemma 4 E4B + genomic model BOTANIC-1 pinpoints causal melon yield mutation in one afternoon — remilouf · 2026-10-14
- 50 practical tips for working with AI: context engineering, voice, and full toolchain control — koltregaskes · 2026-10-14
- TV Time's shutdown cost him 15 years of data, so he rebuilt a 10,000-item media tracker in Obsidian — dSebastien · 2026-10-14
- His Obsidian media system stamps dates automatically as you drag Kanban cards — dSebastien · 2026-10-14
- WIRED Hands-On: Friend 2 AI Necklace Is Louder and Harder to Silence — nordicinst · 2026-10-14
- AI search in one line: scooping the internet's bucket into little answer bubbles — gaganghotra_ · 2026-10-14