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.
More from coding & agent
- ynab-mcp: 49 tools expose the full YNAB API to LLMs for natural-language budgeting — modelcontextprotocol · 2026-09-16
- DaedalMap's MCP connector feeds LLMs CIA World Factbook indicators, 2002-2026 — modelcontextprotocol · 2026-09-16
- OpenAI reportedly prepping Codex Replay to run and compare historical task threads in parallel — testingcatalog · 2026-09-16
- LangChain publishes guide on building custom agent harnesses for domain-specific tasks — LangChain · 2026-09-16
- Building an ERP Q&A bot: the hard part wasn't the model, it was defining "open deal" — Material-Link9151 · 2026-09-16
- LangChain: context engineering is the biggest agent challenge, Deep Agents has it built in — LangChain · 2026-09-16