Chrome team cuts coding-agent vector search from 507 MB to a 7-line loop at 30µs
gaganghotra_ · x · 2026-11-13
Chrome for Developers' Modern Web Guidance 1.0 post details rebuilding its local retrieval engine three times in six months: from 507 MB of C++ bindings to 23 MB of TensorFlow.js, down to a 7 MB package where query embedding is a 7-line loop.
Key design: agentic RAG with progressive disclosure.
- 200+ HTML/CSS/Web API guides (1.7 MB Markdown) can't go into AGENTS.md; the agent only sees a 205-token skill description following Anthropic's trigger guidelines, calling a search CLI on demand.
- On a held-out benchmark of 5,029 real coding-agent searches, Hit@K went 78.2% (whole doc) → 84.0% (heading chunks) → 80.6% (strict 0.38 gate) → 86.1% (+BM25 hybrid) → 93.5% (Top-10, 530-token budget) → 95.4% (Static Int8 7-line loop, 30µs embed), vs 96.9% for a Gemini API cloud reference.
Takeaway: small knowledge bases don't need heavy vector DBs — lexical + static quantization gives zero-dependency, microsecond search near cloud accuracy.
More from coding & agent
- kalomaze: agent memory files misfire—failure notes confuse unrelated contexts instead of helping — kalomaze · 2026-11-13
- Designers urged to try animating with Opus 5.5: let agents iterate from comments, fine-tune keyframes — ycombinator · 2026-11-13
- Agent slop moves up the stack: models rewrite existing library code and build worse replacements — a1zhang · 2026-11-13
- Dev tops AI usage leaderboard: 125B tokens and $30,005 spent in 30 days — jasonkneen · 2026-11-13
- Claude Opus agents reportedly form global hiveminds in under a minute — teortaxesTex · 2026-11-13
- Superconductor adds a private assistant agent that spans all workspace projects — sergeykarayev · 2026-11-13