DigUp Uses EmbeddingGemma 2 for Local Mac Search

DigUp is a native Swift open-source macOS menu bar app, published under the MIT license, that provides local multimodal search without cloud dependencies. The utility integrates Google DeepMind's EmbeddingGemma 2, a 740M parameter model under Apache 2.0, allowing Apple Silicon systems running macOS 14 or later to search files via Shift + Command + Space. Everything operates on-device using an 8-bit quantized build of the model via llama.cpp with Metal acceleration.

Content ingestion combines Apple's ImageIO, PDFKit, AVFoundation, and Vision frameworks—with OCR backing roughly 25 languages—and stores vectors in a single SQLite database. Encoding operates under strict budgets: 280 tokens per image, 140 per video frame, 25 tokens per second of audio, and an 8,000-token cap for text. Audio files are processed in 5.5-minute passes using 30-second overlapping windows, while video is sampled once per shot. A shared 768-dimensional vector space enables unified cosine similarity ranking and keyword matching across media types and code.

After a single 865 MB initial model download, DigUp collects zero telemetry and runs entirely offline. Resource usage remains low because the embedding process exits after work, loading a 250 MB text encoder only during active queries; Google notes corresponding mobile memory baselines of 191 MB for quantized text and 567 MB for the full model on a Pixel 11 Pro. Initial indexing scales from minutes for hundreds of files to roughly an hour for large folders, with a llama.cpp-sized codebase indexing in 20 minutes and adding 130 MB of storage before offering near-instant updates.

For software engineers and AI practitioners, DigUp serves as an efficient reference implementation for local retrieval-augmented generation systems across more than 100 languages. By consolidating deep links to specific video frames, timestamps, PDF pages, and code lines into a lightweight interface, it demonstrates how low-footprint multimodal models can enhance everyday workstation search without data leakage.