mertmetin/searchbench
Python
Public
README.md
2014 / 2026
searchbench
Product search, 2014 vs 2026, measured
My 2014 Django + Elasticsearch price-comparison search, rebuilt to answer one question properly: how much better does modern search actually get, and how do you know? Keyword, semantic and hybrid retrieval, scored against Wayfair's WANDS relevance labels.
## highlights
- Hybrid search (BM25 + embeddings, reciprocal rank fusion) beats the 2014 approach on 217 of 480 queries
- An off-the-shelf reranker made results slightly worse at 20× the latency, so it isn't the default
- FastAPI web app with facets and a side-by-side 2014 vs 2026 view showing the human judgements