the logo wall got a semantic search toggle and i was sure it worked. flip it off, the button changes colour, everything looks right. then somebody asked why turning it off still does semantic searches, and he was right, the switch was never read. the query code asked a different question entirely, it asked whether the model was loaded, not whether the switch was on. the first time the model finished loading, every search from then on used vectors, switch or no switch. turning it off only repainted the button.
the fix was one line of intent, read the actual switch. and once the behaviour matched the mental model it cleaned up nicely. the logos are embedded once at build time, one file with roughly thirty eight hundred rows of int8 vectors, and the browser never embeds a logo again. with search on, the client embeds the query with a small cached model and does one pass of dot products over the precomputed rows, caption words land first and vectors replace them as the model warms up. with search off there is no model download, no vector file download, no vector maths, just substring matching over the text that already ships inside the data file.
verified live with the same query both ways. off gave eighteen marks from words, zero fetches, instant. on gave ninety marks from semantics and fetched both the model and the vectors. flipped back off with the query still typed, eighteen marks again instantly. it also re-runs the current query when you flip the switch so you never retype, and the choice survives a reload. the lesson is boring and permanent, a setting that looks like it works but reads the wrong thing is worse than no setting at all, because the button lies to you while the code does whatever it wants.