← all posts site ↗

compound vibes need per concept scoring

the font finder scored a whole query as one embedding, and that drowned everything. type something with five separate concepts in it and the single vector averaged them all into a blur that matched nothing well. the fix is per concept scoring, where the query breaks into single words and adjacent pairs, each concept gets its own embedding, and every font is scored by its single best concept match.

so a query about worn metal, asian lettering, and outer space now spikes the distressed families hard on the metal word, the space families on space, and the script families on the asian one, instead of one muddy average. the most relevant stuff floats to the top per facet, and the second source of fonts competes per facet too.

the pipeline blends visual and text similarity against each font's preview, with a full text search and an llm category boost as fallbacks. the only blind spot left is that the preview images are latin alphabet specimens, so they can't show asian script in the visual leg. the answer is honest about it: a plain single word query still returns the same solid set as before, and the harder the query gets the better the per facet behavior is.