Named-Entity Recognition (NER) for ecommerce search isn’t the academic sport version (people, places, organizations); it’s a domain-specific tagger that picks out catalog-relevant attributes. “red nike running shoe size 10” → brand=nike, color=red, type=running shoe, size=10.
The win is huge: detected entities can be mapped directly to facet filters, bypassing the lossy lexical-match step entirely. The remaining unmatched tokens go through normal search. Stores that ship NER often see double-digit lifts on attribute-heavy queries, with no ranking-model changes.
Build it from your own data, not a generic model. Mine your product catalog for the actual brand list, color values, size codes; train a lightweight classifier (or a fine-tuned spaCy/BERT) on a few thousand labeled queries from your own search logs. General-purpose NER models will miss store-specific brands and over-tag noise.