Evidence-Aware Large Language Model Framework for Explainable Root Cause Analysis in Dynamic Cloud-Native Systems under Partial Observability

Beschreibung

Cloud-native systems have also brought a lot of complexity to the operations with distributed architectures, dynamic dependency on services and incomplete observability. Conventional root cause analysis (RCA) methods fail to deliver effective and understandable diagnoses due to using single monitoring signals and pattern guidance. The proposed research suggests an Evidence-Grounded Large Language Model (LLM) RCA framework that combines evidence of multiple sources of operation, such as logs, metrics, traces, deployment events and service dependencies, to produce credible failure explanations. The model uses evidence retrieval, contextual reasoning, hypothesis validation and uncertainty assessment to minimize unsubstantiated diagnostic results. Machine learning based RCA models are experimentally evaluated and gave good results with Random Forest achieving 98.37% accuracy and XGBoost giving 98.25% accuracy. The results affirm that evidence-based reasoning enhances automated diagnosis of incident and transparency and reliability of cloud-native settings under partial observability.