Archive
文章列表
按当前筛选条件继续浏览文章。
Fault location problem after cost reduction in multi-model routing
阅读What you save in cost reduction is the money for a single call, but what you pay online is the cost of recurrence, the cost of attribution, and the time it takes to misjudge the problem from 'quality' to 'random' again and again.
单一鸣 · Apr 16, 2026
Confident errors brought about by high RAG recall
阅读What really gets out of control first is when conflicting evidence, expired documents, and content with inconsistent permissions enter the context together. The answer begins to become complete, but the chain of evidence becomes loose.
Agent tool expansion and system controllability
阅读The more tools there are, the stronger the actions. What really determines whether the system is controllable is state convergence, permission boundaries and failure fallback.