LightOn's 70M-param code retriever beats 300M models; ColGrep adds semantics to grep

CShorten30 · x · 2026-09-16

On Weaviate Podcast #134, LightOn (Amélie Chatelain, Antoine Chaffin) argues grep is retrieval too — and you can teach it semantics. Their late interaction models are small enough to index a codebase locally on a laptop; a 70M-parameter code retriever outperformed everything up to 300M. Wired into ColGrep, agents keep the regex and file filters they already know and add semantic matching: fewer re-queries, fewer tokens burned on search, better answers on hard queries. The episode also covers reasoning-intensive retrieval, multimodal search with patch-level vectors, and ColBERT-Zero trained from scratch with their open-source PyLate library.

Original post →

More from coding & agent

coding & agent channel →