← all posts site ↗

searching my saved posts by vibe not keyword

asked for an engine that searches the things i've saved by style instead of text. the first pass was keyword recall and it worked fine for recall, ask for a futuristic game setting and it surfaces all the cyberpunk concept art i've collected. but the real ask was adjacency. i wanted to say give me a summer fit and have it surface a watch nobody ever tagged as summer, purely because its photo lives in the same visual neighborhood as my linen fits.

keyword can't do that because text has no concept of two things sharing a palette. the fix is visual. mymind happens to expose a preview image url for every saved card, so i dumped the whole library to a sqlite file, pulled down the images, and embedded all of them with a clip model running locally, no api credits. thirty five hundred images done in about seventeen minutes. the pipeline that came out of it: expand an ask into a pile of search terms, find the posts the terms overlap on, average their image vectors into one style centroid, then cosine-match the whole library against it.

the proof was space horror, which wasn't a tag anyone had hit. the engine returned posts titled lone figure before a massive alien sphere and the rings at the edge of the world, zero shared keywords, but they read as space horror when you actually look. i verified the top matches by eye instead of trusting the math, then built the context bundle: title, tags, description, the dominant colors pulled from the real pixels, and the preview path. that's what a model gets to build a style from, not a guess, my actual saved aesthetic.