Live, against the canon

Identify a room

Show it a photograph of a room and it tells you which of the 197 rooms in this vocabulary it most resembles, with the runners-up and how close each one came.

How well it works

Photograph to room55.6%

Written description to room50.0%

Photograph to style15.0%

Named correctly on first guess, over 180 held-out photographs against 197 candidates. The photograph names the room better than the words written to sell it.

Why it will not name a style

Because it cannot — the third bar is a guess with a decimal point on it. The index this endpoint ships carries space vectors and no style vectors at all, so the refusal is in the data rather than in this paragraph, and editing the paragraph would not change it.

What happens to it

The image is sent once, embedded once, and discarded. Nothing is written to disk, nothing is logged, and there is no account to attach it to. The model is voyage-multimodal-3.5, which puts pictures and text in one space; the term vectors were computed once and ship with the endpoint.