标签: AI

生产前如何评估LLM?

生产前如何评估LLM?

GitHub 团队在将 LLM 用于 secret scanning 时发现,基准测试表现好不等于生产环境可靠,并据此总结出一套以产品决策为导向、离线评估当作集成测试的大模型上线前评估方法。

UI locks text box with “Context Window Exceeded” at ~500k tokens despite 1M context model limit

UI locks text box with "Context Window Exceeded" at ~500k tokens despite 1M context model limit

该报错并非前端本地 token 计数触发的误拦,而是模型服务端真实返回“上下文窗口超限”错误后,LobeChat 的 UI 将该错误渲染到输入框并锁死输入。优先排查方向是:当令牌用量接近模型上限时,先点击错误提示中的 Compact(压缩历史)按钮,或尝试将输入栏模式从 Agent 切到 Chat

Keenable, which is building a web search index for AI agents and says several AI labs are using its API, raised a $26M seed led by Accel (Anna Heim/TechCrunch)

Keenable, which is building a web search index for AI agents and says several AI labs are using its API, raised a $26M seed led by Accel (Anna Heim/TechCrunch)

为 AI 代理构建专用网络搜索索引的 Keenable 宣布完成 2600 万美元种子轮融资,由 Accel 领投,并透露已有多个 AI 实验室在使用其 API。这笔投资说明,在 AI 代理大规模调用外部实时信息的趋势下,搜索基础设施正成为新的竞争焦点。