📄 中文摘要
代理 RAG 系统在生产环境中可能会出现无声失败，导致资源浪费和性能下降。主要的失败模式包括检索混乱（Retrieval Thrash）、工具风暴（Tool Storms）和上下文膨胀（Context Bloat）。检索混乱指的是系统在多个检索请求之间频繁切换，导致效率降低；工具风暴则是指系统在处理任务时过度依赖多种工具，增加了复杂性和错误风险；上下文膨胀则是指系统在处理信息时积累过多无关上下文，影响决策质量。识别这些问题的早期迹象对于优化系统性能和控制云计算成本至关重要。
代理 RAG 失败模式：检索混乱、工具风暴与上下文膨胀（及早识别的方法）
📄 English Summary
Agentic RAG Failure Modes: Retrieval Thrash, Tool Storms, and Context Bloat (and How to Spot Them Early)
Agentic RAG systems can experience silent failures in production, leading to resource waste and performance degradation. Key failure modes include Retrieval Thrash, Tool Storms, and Context Bloat. Retrieval Thrash occurs when the system frequently switches between multiple retrieval requests, reducing efficiency. Tool Storms refer to the excessive reliance on various tools during task processing, increasing complexity and the risk of errors. Context Bloat involves the accumulation of excessive irrelevant context during information processing, negatively impacting decision quality. Early detection of these issues is crucial for optimizing system performance and managing cloud computing costs.