New Local Search Application Utilizes Google DeepMind EmbeddingGemma 2

A new open-source application, DigUp, has been released for macOS, enabling users to perform local searches across diverse file types using Google DeepMind’s EmbeddingGemma 2 model. According to reports on Reddit, the application allows for semantic retrieval, meaning users can locate specific content within files using natural language descriptions rather than exact keyword matches.

Technical Specifications and Performance

The application is built as a native Swift app for Apple Silicon running macOS 14 or later. According to the developer’s documentation on GitHub, the software utilizes the ggml-org Q8_0 GGUF model (approximately 865 MB) via llama.cpp with Metal acceleration. Notably, the application does not require a Python environment to operate.

The developer reports that the search function only loads the text encoder, which occupies approximately 250 MB of memory. Performance metrics provided by the developer indicate that search results appear roughly 0.1 seconds after a user finishes typing. During the indexing process, memory usage is reported to peak under 2 GB. For multimedia files, the application processes audio and video by utilizing 30-second windows and one frame per shot to facilitate searching.

Multimodal Search Capabilities

The software is designed to index and retrieve information across several formats, including text, images, audio, and video. Based on developer claims, the application supports the following functionalities:

  • Video and Audio: Users can search for specific visual or auditory elements, such as “zebra in a video” or “where they talk about sleep,” to jump to the relevant timestamp.
  • Documents: The tool can locate specific clauses within PDF files and highlight the relevant text.
  • Images: The application supports multilingual searches, allowing users to find images using languages such as Bengali or Arabic.
  • Code: When code search is enabled, the application can locate specific functions within a user’s editor using snippets like code: retry with backoff.

Privacy and Deployment

The application is released under the MIT license and is free to use. A key feature highlighted by the developer is that all processing occurs locally on the user’s hardware. The software only requires an internet connection for the initial model download and periodic update checks, an option which users can disable.

While the official release is for macOS, a Windows version referred to as “VectorDash” has been mentioned in community discussions on Reddit. According to user comments, this version allows for execution on either the CPU or GPU, though CPU execution may result in slower performance on lower-specification machines.

As with many software releases hosted on GitHub, these performance metrics and capabilities are provided by the developer and have not been independently verified by third-party benchmarks.