← all posts site ↗

fixing taste engine groups by measuring

the groups feature kept mixing up totally unrelated saved posts and no threshold fixed it, so i finally measured instead of guessing. three separate causes surfaced. every imported screenshot shared the same title, screenshot, date, platform, so hundreds of rows had one identical text vector. a screenshot image embedding describes the phone ui, not the post inside it. and clip text scores saturate, two unrelated screenshots scored point seven to point eight five similarity, so no threshold could ever separate them.

the fix was to stop using clip for screenshots entirely. they now get a hashed bag of words over the ocr text plus the explainer, with boilerplate stripped, which scores unrelated text near zero and identical text at one. ocr now runs on every tall screenshot at boot, and the grab bag of sixty nine items collapsed into groups with a real topic: gamedev, workspace, cyberpunk art. three hundred and eight groups live and the model chat tools still answer.

the lesson is the same as always: when similarity smells wrong, look at the vectors, not the threshold. what looked like one bug was three compounding causes, and one model swap fixed the whole thing.