DigUp Uses EmbeddingGemma 2 for Local Mac Search

A new MIT-licensed Mac menu bar application named DigUp provides local semantic search across user files on Apple Silicon hardware running macOS 14 or higher. Powered by Google’s Apache 2.0-licensed EmbeddingGemma 2 model, the tool operates entirely on-device using llama.cpp on Metal following a one-time 865 MB download. The app maps photos, PDFs, text documents, audio clips, videos, and source code into a shared 768-dimensional vector space.

DigUp processes local content by converting each item into a list of 768 numerical values representing its underlying meaning, bypassing the need to generate preliminary text captions. Videos are broken down into distinct visual frames and audio components before being embedded across overlapping windows. The application supports semantic queries in over 100 languages and provides optical character recognition for roughly 25 languages through Apple Vision. Search indexes are stored locally in a single SQLite database alongside a traditional keyword index for literal matches.

By leveraging Google’s 740-million-parameter embedding architecture, DigUp delivers millisecond-scale query performance for complex cross-modal retrieval. Users can type a natural language prompt to locate specific video moments, image elements, or document clauses in parallel. For privacy-conscious developers and technical professionals, this implementation demonstrates how high-dimensional multimodal retrieval can run efficiently on local desktop silicon without relying on remote server infrastructure.