Smart Albums

A smart album is defined by a predicate rather than a membership list: it names a query, and its contents are whatever currently matches. An asset that starts matching appears in it; an edit that takes an asset out of the predicate removes it. Nothing is copied and no membership row is written, which is why the same result is reachable by running the query directly.

Because the server holds only ciphertext, every predicate is evaluated on your own device, against the local index. A smart album is a saved query over a library only your devices can read, not a view the server materializes.

What a predicate can say

The grammar, its closed field and operator sets, and how two devices converge on a concurrently-edited definition are owned by Organization — System / Smart Albums Views. Definitions travel in the library-settings document rather than in a per-asset sidecar, because they belong to the library rather than to any one asset.

The kinds that depend on machine learning

Several album kinds people expect — one per person, trips clustered by place and time, pets, food, scenes, a best-of-year selection — are predicates over tags and embeddings that an on-device model produces. Those models, what each is allowed to assert, and how an embedding records the model version that made it are owned by AI/ML Integrations.

That doc also carries the constraint that shapes them: every device must be able to run the same model, or a library’s organization depends on which phone happened to import a photo. Model choice is bounded by the weakest supported device, not the strongest.