Your iPhone Photos app, but it understands 'brown dog on bench'

What it does

Queryable is the open-source iOS app that indexes your local Photos library using Apple’s MobileCLIP model. It embeds every image into a vector, then compares that against a text embedding from your natural-language query—say, a brown dog sitting on a bench—to surface the top matches. The entire pipeline runs on-device via Core ML, so your photos never leave the phone.

The interesting bit

The project splits CLIP’s dual-encoder architecture into separate Core ML models, letting the image encoder chew through your album once while the text encoder answers queries on demand. The author freely admits they are not a professional iOS engineer, which gives the whole thing an endearing “weekend project that accidentally shipped” energy.

Key highlights

• Runs fully offline; no network calls to Apple, Google, or anyone else
• Uses Apple’s MobileCLIP s2 model by default, trading off efficiency and precision
• Supports free-form natural-language queries instead of rigid category tags
• Ships with Jupyter notebooks showing how to split and export PyTorch CLIP/MobileCLIP into separate Core ML encoders
• Spawned an Android port (PicQuery) and inspired a closed-source macOS variant (Searchable)

Caveats

• The author warns the Swift code is “poor” and suggests focusing on the model loading and computation parts
• Core ML export of the ImageEncoder still carries precision errors that may need normalization tweaks

Verdict

Good for developers prototyping on-device semantic search or Core ML model splitting. Not the right fit if you need a polished, production-ready photo manager.