Semantic code search vs grep: a measured test on the requests repository

A developer benchmarked semantic embedding search against grep on psf/requests v2.32.3 using ten questions where the obvious keyword returns nothing. Embeddings alone placed the correct chunk in the top five for four of ten queries, while a simple keyword ranker found six. Combining both approaches reached seven in the top ten, rising to nine when tests and documentation stopped competing with source code. The pipeline, a Python rewrite of a desktop app's indexing logic, ran on an 8 GB M1 Mac using CPU only. The author also notes that code tokenizes more heavily than prose, so the usual four-characters-per-token approximation underestimates chunk size.