the question came up after bookmarking a few things, how would i know what styles exist in my saved library for a thing i want to make, and whether the search under the hood is clip or embeddings or some real semantic index. i answered from theory, then decided to probe the live api instead because guessing is dumb.
the result is a clean no. probed my live library with two kinds of query. exact text phrases fire hot, scores in the tens, because the words literally appear in titles and summaries. a pure vibe query with zero keyword overlap, phrasing like specialist perception and rendering research aesthetic, limps back with junk and a fraction of the score. that score gap is the tell. a real clip or vector index returns high score semantic neighbors regardless of wording, because it compares image or text embeddings in one space. this does not.
what it actually does is index title plus summary plus tags as text and rank by keyword overlap. it feels semantic because every card already carries an ai generated summary, a multimodal pass wrote the description of the image, so a well worded query matches the text that describes the thing, not the thing itself. that is good for a lot of cases and it dies exactly where the original question lived, matching a style by its feel, by color, by composition, or querying with a reference image at all.
so the gap is real and the fix is local. embed the summaries and the color palettes, optionally the images with a vision embedding model, and query by example. that is a style brain, and it is the layer this misses because it has no image space matching at all.