after semantic search landed on the logo wall the obvious question came up, can this find marks by how they look, not just by the words near them? ask for a ball and you want ball-shaped things plus anything whose name says ball. the honest answer was no, the old search used caption text plus a small auto-tagger, which is a proxy for shape, not shape. the real vectors were sitting on disk the whole time, clip image embeddings from the pixels, computed weeks ago for another project. nobody had wired them to the wall.
so the wall ended up with two vector spaces, both made once at build time. one is the text embedding on caption, client, industry and tags, which answers what the description means. the other is one row per mark from the actual pixels, quantised to int8, about two megabytes, which answers what the mark looks like. a shape query costs one trip to the clip text tower, cached per query, then roughly thirty eight hundred by five hundred and twelve int8 dot products in the browser. the maths stays local, the server just answers a single vector request.
the results blend three signals, shape gets half the weight, meaning about a quarter, caption words the rest, so words land first and shape and meaning replace the set while the model warms. verified with real queries. ball returns tennis balls, globes, aperture marks, things that are actually round. cow returns horns and bull heads. star and arrow do what they say. with search off it is words only, no vector file download, no extra requests, instant.