Clio compares grep ranking to RAG index in Claude Fable 5.2
The author tested one-shot file search performance for AI coding agents on 40 real queries, comparing vector search, grep variants, and a RAG index. The tests showed that ranking grep hits by BM25 or distinct question words significantly outperformed basic match counts and raw vector search, even tying or beating the unmaintained RAG index. The author suggests scoring local grep rankers on custom question sets before building complex knowledge base indices. Open article →