CoreAIKit

PhotoSearch

Semantic photo search, fully on-device: the photo library is indexed with CLIP image embeddings (Neural Engine, ~4 ms per photo), and free-text queries like “red bike at the beach” rank photos by cosine similarity.

Built on CoreAIKitVision — the whole ML surface of this app is encoder.encode(image:) / encoder.encode(text:).

Run

xcodegen generate
open PhotoSearch.xcodeproj

Run on an iPhone (photo library + Neural Engine). First launch downloads the CLIP bundle (~290 MB) from the Hugging Face Hub, then indexes the library once; embeddings persist under Application Support, so later launches only index new photos.

Notes

• Brute-force search (one dot product per photo via vDSP) — instant up to tens of thousands of photos; no vector database needed.
• Index storage is a flat Float32 file + an id list, ~2 KB per photo.

This site is open source. Improve this page.