BM25 matches 8B embedding model for tool retrieval

An evaluation of tool-retrieval methods for AI agents tested performance across tool catalogues ranging from 1,000 to 3,794 items. The author found that a simple BM25 keyword search matched the performance of an embedding model 24 times its size while running significantly faster, and that providing full tool documentation improved retrieval more effectively than upgrading to a larger model. Open article →