DigUp: Open-Source Mac App for Local Multimodal File Search via EmbeddingGemma 2
Developer A-Rahim released DigUp, a free, MIT-licensed native macOS application that runs Google DeepMind's EmbeddingGemma 2 model locally to search personal files using natural language. Built with Swift, llama.cpp, and Metal GPU acceleration, the app enables fast semantic search across text, images, audio, video, and code across multiple languages. DigUp demonstrates how lightweight, quantized multimodal embedding models can deliver instant, privacy-focused desktop search without relying on cloud APIs or heavy Python runtimes. It makes local semantic search practically accessible for Apple Silicon Mac users with minimal resource usage. The app downloads an 865 MB Q8_0 GGUF quantized model and requires only around 250 MB of memory for search queries, returning results in about 100 milliseconds. Indexing processes audio in 30-second windows and video frame-by-frame, keeping peak memory consumption under 2 GB on macOS 14+ devices.
## BACKGROUND
Multimodal embedding models map text, images, and audio into a shared mathematical vector space, enabling users to search media files based on descriptive meaning rather than exact filenames. GGUF is a binary file format created by the llama.cpp project that allows quantized local AI models to execute efficiently on consumer hardware using acceleration frameworks like Apple's Metal.