#1Show HN: Semble – Code Search for Agents Using 98% Fewer Tokens
Semble is a new code search library purpose-built for AI agents. Instead of grep-and-read, agents query in natural language and get back only relevant code chunks, scored via a fusion of semantic embeddings (Model2Vec) and lexical matching (BM25). It indexes repos in ~250ms, answers queries in ~1.5ms on CPU alone, and claims 99% of the retrieval quality of larger transformer models — all while slashing token consumption by 98%.