Search Relevance Lab
BM25, sentence embeddings, open WANDS dataset
Search teams argue about relevance from a handful of example queries. This app runs one query through keyword, semantic and re-ranked search on a catalogue of 42,994 products, and scores each method against labels that people made by hand.
Letting any query word match lifted nDCG@10 from 0.406 to 0.704. The learned blend added 0.034 on top.
Semantic search is under-scored, because only 65% of its top results have a label.
Open WANDS dataset, 192 held-out test queries, scored 16 Nov 2026. Furniture and home goods only. Free hosting can take a minute to wake up.