DigUp is not a conversational assistant. It is a local search engine that uses Google DeepMind's new multimodal embedding model to index personal files. The app is free and open source under the MIT license, targets Apple Silicon with macOS 14 or later, and runs entirely on device. EmbeddingGemma 2 ships as a Q8_0 GGUF file of 865 MB, downloaded once, and is executed through llama.cpp with Metal inside a native Swift app. There is no Python in the runtime path, which keeps the dependency footprint small.

Search loads only the text encoder, roughly 250 MB, and returns results about a tenth of a second after the user stops typing. Indexing peaks below 2 GB of memory and the helper process exits when the job is done. For audio and video, DigUp splits content into 30-second windows and captures one frame per shot. This does not preserve every instant, but it creates enough anchor points for queries such as zebra in a video to open the clip at the right moment, or the clause about pets in the lease to show the PDF page with the words highlighted. The same query in Bengali or Arabic finds the photo of a dog on the beach: the model aligns text and images in a shared space, so query language is not a rigid filter.

Code search adds another layer. With the mode enabled, code: retry with backoff opens the relevant function in the editor. That points to a broader shift: indexing code behavior, not just documents, can reduce reliance on separate code search tools in on-premise systems.

The main constraint is not the model download but index construction. The design choices—loading only the text encoder, exiting the helper after indexing, keeping peak memory under 2 GB—show an effort to limit RAM and storage pressure. For teams evaluating on-premise deployment, this is a concrete reminder that a quantized GGUF and a Metal-enabled runtime on Apple silicon can keep a multimodal local index within the limits of a single laptop. No dedicated server or cluster is required. That changes incentives relative to cloud search services: files stay on the machine, and the app goes online only for the model download and an update check that can be turned off. Data sovereignty becomes an attribute of the deployment rather than an abstract requirement. AI-RADAR offers analytical frameworks at /llm-onpremise to compare these trade-offs.

Q8_0 is not FP16: quantization introduces approximation in the weights. But for personal content retrieval, the benefit is running the model without exhausting memory. The project does not claim to replace a distributed vector database; it shows that a multimodal embedding pipeline can live inside a native, signed and notarized app with minimal entry cost. Structurally, this signals that local embedding-based search is moving beyond prototypes and becoming relevant for those who do not want to move metadata and content outside their device.