Hover any product card for its full details — category, rating, score, and why it was shown.
Six strategies, and how they differ
Every strategy receives the exact same baseline candidates — a widened pool from search-service
(hybrid dense + lexical retrieval), narrowed by a hard eligibility filter before any strategy runs — and either
leaves them alone, re-ranks them, or adds/removes products — never a from-scratch retrieval of its own.
Numbers below are from RESULTS.md's independent WANDS relevance eval — scored against WANDS's own human
relevance grades, which no strategy is trained or tuned against, rather than the clickstream-derived
labels the strategies themselves read from
(full methodology, the second implicit-clickstream table, and honest caveats in RESULTS.md).
Click any card for how it works and its grounding in the literature.