BM25 is no longer a clear winner in 2023?

Table of Contents

Very nice paper https://lnkd.in/eyfhuyAi from Metarank Labs

Similar to Vespa papers, Metarank papers are based on hard benchmark figures, which is the best/only way to understand facts.

— Summary —
> BM25 was a tough contender 2 years ago, when the BEIR benchmark and SBERT models reigned supreme.
(Probably explaining why Elasticsearch is now fully commited to vector search)

> Nowadays, new MTEB benchmark https://lnkd.in/exGBk3ND is dominared by new models like Microsoft E5.

The trend is that keeping your current vector search and waiting for model embeddings improvement is a winning strategy !

WPSOLR + WooCommerce + Weaviate + SBERT embeddings (waiting for E5?): https://wpsolr.com

#wpsolr #bert #sbert #wordpress #woocommerce #huggingface #weaviate #metarank

Read more related content