An Intelligent AIOps Framework Based on Multi-Agent Collaboration and Dual-Context Retrieval

Enterprise operations and maintenance knowledge and runtime evidence are often scattered across manuals, incident reports, monitoring records, and emergency plans. Conventional keyword retrieval is weak in semantic matching and evidence localization, while a single question-Answering model often struggles to integrate heterogeneous operational signals for fault-Assisted diagnosis. To address these issues, this paper proposes an intelligent AIOps framework based on multi-Agent collaboration and dual-context retrieval. The framework first performs document-level global recall to preserve scenario background and then chunk-level local ranking to identify actionable evidence. A unified orchestrator coordinates knowledge, monitoring, decision, notification, and host-management agents to interpret tasks, invoke bounded skills, aggregate cross-source evidence, and return traceable results. A FastAPI and React prototype is implemented and evaluated on a reproducible benchmark built from the open-source Rootly logs dataset and project runbooks. On 85 retrieval cases, dual-context retrieval reaches Recall@5 of 0.98, improves MRR from 0.90 to 0.92, and reduces average retrieval latency among hybrid methods. On 94 fault-diagnosis cases, the full workflow improves Top-1 accuracy from 0.57 to 0.79, action correctness from 0.68 to 0.86 and evidence completeness from 0.47 to 0.91. These results provide preliminary, reproducible evidence that the proposed framework improves evidence grounding and traceable diagnosis in a practical prototype setting.