Hybrid Search

Anserini, an open-source toolkit for reproducible information-retrieval research, publishes an experiment that should be required reading for anyone about to ship hybrid search. Its authors take two retrievers, BM25 and the bge-base-en-v1.5 embedding model, run both on the 18 datasets of the BEIR benchmark, and fuse the two result lists three ways. Adding the raw scores gives an average nDCG@10 of 0.440, barely above BM25 alone at 0.424 and well below the embedding model alone at 0.508. Rescaling each list to the range 0 to 1 before adding gives 0.531. Fusing by rank position alone, with no scores at all, gives 0.516.