jevPOST
Jev 代理循环里只剩两处需要 LLM
开发者把工具返回结果的过滤从 LLM 生成筛选条件改成 Jev 直接做语义过滤,称小数据集下比生成的筛选条件更准。
James Ward
@JamesWard2026.09.23收录原推发布于 2026.09.23
James Ward @JamesWard原推发布于 2026.09.23
For my Jev-based agentic loop there are now only two places I need an LLM as a tool: summarization & parameter parsing. Initially I was using the LLM to create a filter criteria for items returned from tool calls (like how your AI code assistant defines grep & find calls). But now I'm just having Jev do the filtering. It can't create a filter criteria like an LLM, but for small enough data sets, Jev can just do the semantic filter - which is probably more accurate than generated filter criteria.
x.com/jamesward/status/2102468057494917237采集于 2026.09.23 10:15