The GitHub README describes indexing images, PDFs, documents, audio, and video with EmbeddingGemma 2, and opening results at relevant pages or media moments. This matters because the model is not presented only as a text embedding component: the app turns a shared multimodal representation into a local search interface across heterogeneous personal files.
First implication: This makes cross-format retrieval available as a local desktop workflow, but the repository's scope is bounded by folder selection and its stated file-processing rules.
Thesis
Local multimodal search makes embedding models tangible as an interface to a user's mixed-format archive.
Observed facts
The DigUp README says the app indexes selected images, PDFs, documents, audio, and video using EmbeddingGemma 2.
1The README states that DigUp requires Apple Silicon and macOS 14 or later and downloads an 865 MB model on first launch.
1The project says indexed files stay on the Mac and that it does not require an account or use telemetry.
1
Mechanism
DigUp indexes user-selected folders with EmbeddingGemma 2, then matches natural-language queries against representations of text, images, audio, and video; it can also match exact strings.
Why now
the project packages a multimodal model into a free Mac application, rather than requiring users to assemble an indexing pipeline themselves.
Implications
This makes cross-format retrieval available as a local desktop workflow, but the repository's scope is bounded by folder selection and its stated file-processing rules.
The app's utility depends on the quality of indexing and retrieval across different file types, which is not independently assessed by the supplied source.
Evidence boundary
These details come from the project's GitHub README; they establish the developer's stated design and requirements, not independent testing of privacy, recall, speed, or reliability.
Watch conditions
Whether independent users report reliable retrieval across varied personal archives, especially for audio and video, will show how well the described workflow holds in practice.
Whether the stated local processing and no-telemetry properties are independently audited or tested.