Finding similar items (ADR-0017) Diagram

Three signals, one score

Looks: how alike two pictures are, to the centre of the selection; a film's looks are the mean of six frames
Text: TF-IDF over a document's words and any item's summary; it matches words, not meanings
Known facts: the same event, person, pet, place or thing, the same album or folder, a date within three days
Each signal adds a weighted share; the reasons come back as text, such as 'looks alike 0.86 · same album · 2 days apart'

What leaves the PC

Looks vectors, extracted text and the ranking stay in the edge's catalog, like face embeddings
The chat reads the ranking only when the person asks it to group
The cloud learns only the tags a person confirms, through the usual A-Box sync

Before the looks are made

Looks need the 350 MB model once and a pass over the photos, as a job
Until a photo has its vector, it is compared by text and known facts only, and the answer says how many lack looks