BM25

BM25 is a classic ranking algorithm developed around 1994 by Stephen Robertson and Stephen Walker that combines term frequency and inverse document frequency in a probabilistic framework. It has served as the backbone of search engines and Lucene/Elasticsearch for decades; despite its simplicity, it remains competitive with modern dense methods on many tasks. In the RAG world it has found a second life as the sparse half of Hybrid Search, covering exactly the rare-term, code, or ID lookups where Embedding-based retrieval tends to falter.