This is a subtle but (hopefully) significant improvement to search and
related posts
.Before: search and
related posts
relied on term matches
After: search ranks on term matches and semantic similarity while related posts
relies exclusively on semantic similarityThis should improve both evergreen-ness of content and rabbit-holing. IMO
related posts
have improved an incredible amount, but search is also much better at finding relevant matches.For months, @elvismercury hollered about text embeddings, but I didn't know what text embeddings were. This change uses on them by mapping a search query or item to a vector of meaning and looking for items with similar vectors.
what are you working on
match_score*zaprank
but it might make more sense to domatch_score*log10(zaprank)