BMX: A Freshly Baked Take on BM25

Mixedbread AI and the Hong Kong Polytechnic University have introduced BMX, an advanced lexical search algorithm, promising significant improvements over the industry standard BM25. This algorithm leverages entropy-weighted similarity and weighted query augmentation (WQA) to integrate similarity and semantics into the search process. The research demonstrates BMX's superior performance on various benchmarks, including BEIR, BRIGHT, and multilingual datasets. BMX’s reference implementation is available in the open-source Baguetter library.

• BMX outperforms BM25 in multiple benchmarks.
• BM25 and BMX were compared in the BEIR benchmark.
• BMX uses entropy-weighted similarity.
• BMX includes weighted query augmentation.
• The Baguetter library implements BMX.