Jevjitsu: Or, How We Tried Generalized Classifiers on Everything - Qdrant

This post introduces experiments with Jev, a decision model, for several classification tasks within retrieval systems. It details Jev's application in reranking search results, demonstrating that Jev Score and Jev Choice (1 request per query) offer significant nDCG@10 lift over baseline embedding models and comparable or better results than local cross-encoders. It also explores using Jev to guide Maximal Marginal Relevance (MMR) for reducing repetitive answers and increasing product variety, showing that combining Jev with MMR can improve relevance while maintaining diversity. Furthermore, the post describes using Jev for query understanding and taxonomy routing, where Jev categorizes products and scores them against queries to filter or boost results, demonstrating its utility in improving search relevance on datasets like Amazon-C4.