← all posts site ↗

the taste engine groups everything by vibe

the new inspire idea for my visual bookmark app skips the llm entirely. instead of asking a model what is similar, everything i have saved sits in high dimensional space, images and text fused, clustered by similarity. one tap returns a random group of things that are all actually related. tested on the real vault before building: a mutual nearest neighbor method made one four hundred item blob with six hundred singletons, so i went with k means at about four hundred clusters and a cohesion around point eight four.

the vector is a fusion. image embeddings weighted at point seven, text embedding of title and body at point three, batched through onnx. the cluster module runs at boot, refits every so often, and exposes groups through the rest api and the mcp tools. the web side got a panel with two by two cover collages and a shuffle that passes the last ten ids as an exclude list so it keeps giving fresh sets.

the android app got a groups tab with the same shape, one cluster at a time with no labels, a masonry grid, and a bottom filter bar for kind, source, hue, and tag. if nothing in the current cluster survives the filter it jumps to one that does. a degenerate cluster showed up during deploy, one hundred forty five items sharing one identical vector. not a fit bug, duplicate embeddings from an older pipeline pass, and the refit split them into fourteen distinct ones.