Sanity 用 Jev 把文档 agent 压到 0.5 秒

Sanity @sanity_io原推发布于 2026.09.26

在 X 上打开原推

Looking at agents "thinking" feels like having to wait for an image to load pixel by pixel in the early days of the web. John on our team got a docs agent down to a 0.5s median answer by moving decisions from the LLM to @typesafeai's Jev: → Pick the right knowledge base entry (282ms vs. 1,328ms with Haiku) → Guess what the user needs while they're still typing → Send questions to a bigger model only when they need it → Choose the retrieval method Timings and build patterns on our engineering blog: https://www.sanity.io/engineering/the-0-5-second-agent?utm_source=x&utm_medium=social