Why Claude Code still uses grep: a field guide to multi-vector search
antoine_chaffin · x · 2026-07-24
LightOn’s Amélie Chatelain argues that multi-vector / late-interaction search is better suited for agents than single-vector dense retrieval, especially for longer queries, longer contexts, and long documents.
The attached slides frame the core idea:
- Single-vector search compresses each query/document into one embedding.
- Late interaction keeps one representation per token, which can improve generalization and retrieval quality on domains and inputs that were not fine-tuned for.
- The talk explicitly asks why Claude Code still reaches for grep even with strong LLMs and embedding models: lexical search remains useful as a practical tool inside coding agents.
The post links to a full lecture described as a field guide to the space, aimed at people building AI-powered search and agent workflows.
More from coding & agent
- A coding-agent user wants `/goal` to batch tasks and compact between them — mattpocockuk · 2026-07-24
- GPT 5.6 Sol and Luna reportedly built a real-time vision stack over MCP — MoonL88537 · 2026-07-24
- A GitHub repo nearing 200,000 stars reduces coding-agent behavior to four rules — import_jmr · 2026-07-24
- One prompt adds voiceovers to a Three.js game through ElevenLabs — majidmanzarpour · 2026-07-24
- OpenClaw rewritten in Go now runs in 35 MB RAM with a 25 MB binary — bigaiguy · 2026-07-24
- Go rewrite of OpenClaw cuts a multi-agent platform to 25 MB and 35 MB RAM — bigaiguy · 2026-07-24