智能体
x.com ·
@aiDotEngineer
AI Engineer 检索专场笔记：Legora 延迟问题、llms.txt 失效与多尺寸索引
AI Engineer 汇总检索专场要点：Legora 因已完成与进行中项目共享数据库分区，P99 搜索延迟从 100 毫秒升至 20 秒，改为每项目一个 namespace 后解决；约半数被测站点发布 llms.txt，但 agent 几乎不读取，而是直接访问文档页和首页；AI21 以六种 chunk 尺寸索引同一数据并按查询择优，召回较任一固定尺寸提升约 20–40%；agent 有 30–50% 的 token 用于检索，SID.ai 的 RL 训练检索模型据称比前沿模型快约 20 倍、成本约百分之一；agent 在合同中搜索“30 days”时无法区分截止期、宽限期与保留规则。
AI推荐理由 汇总了 AI Engineer World's Fair 2026 检索专场中的多条具体工程经验：Legora 因共享数据库分区导致 P99 延迟从 100ms 升至 20s、按项目分 namespace 解决；约半数站点发布 llms.txt 但 agent 几乎不读；AI21 多 chunk 尺寸索引使召回提升约 20–40%；agent 30–50% token 花在检索上；SID.ai 的 RL 检索模型据称比前沿模型快约 20 倍、成本约 1/100。信息密度高、有具体数字和场景，但均为转述，未经独立验证。
82
浏览 428
点赞 1
评论 1
转发 2
正文
Notes from classroom:
- Legora's search latency went from 100 milliseconds to twenty seconds at the 99th percentile once finished legal projects and live ones shared database partitions. One namespace per project fixed it.
- About half the sites Liad Yosef's team tested publish an llms.txt file for agents. Almost none of the agents they ran read it. They went to the docs page, then the homepage.
- AI21 indexed the same data at six chunk sizes and let hindsight pick the best size per query. Recall improved something like 20 to 40 percent over any single fixed size, so now they index at several sizes at once and merge the rankings.
- Agents spend 30 to 50 percent of their tokens on search. [SID.ai](http://SID.ai)'s RL trained search model does the task about twenty times faster than a frontier model at roughly a hundredth of the cost.
- An agent searching a contract for "30 days" cannot tell a deadline from a grace period from a retention rule. Code, full of exact names to grep for, was the easy case.
Search & Retrieval playlist:
[youtube.com/watch?v=IN-rb-…](https://www.youtube.com/watch?v=IN-rb-9WmiY&list=PLcfpQ4tk2k0W3T87n_MZGaV9WfWOmEWtQ)