Operational Knowledge Graph Architecture for Explainable Artificial Intelligence in Cloud-Native DevOps Operations

Abstract

Modern artificial intelligence for IT operations (AIOps) platforms increasingly rely on large language models to assist with incident investigation, deployment reasoning, and operational decision support in cloud-native environments. Most existing systems, however, treat infrastructure telemetry, container orchestration state, continuous integration and deployment (CI/CD) metadata, monitoring signals, and historical incident records as disconnected data sources. This fragmentation limits the explainability of AI-generated recommendations and increases the risk of hallucinated or unsupported conclusions. The primary objective of this study is to design an Operational Knowledge Graph (OKG) architecture that unifies these heterogeneous data sources into a single semantic model capable of grounding large language model reasoning in verifiable infrastructure relationships. A design science research approach was followed, combining a structured review of knowledge graph-based root cause analysis and graph-grounded retrieval-augmented generation (GraphRAG) literature with layered architectural decomposition informed by operational experience with production Kubernetes and cloud infrastructure. The methodology produces a five-layer architecture spanning data ingestion, semantic modeling, temporal graph construction, GraphRAG-based retrieval and reasoning, and explainability. The architecture is demonstrated through two representative operational scenarios drawn from common cloud-native failure classes: a container image pull failure caused by a missing network path between private compute nodes and a container registry, and a load balancer provisioning failure blocked by an account-level policy boundary. In both scenarios, graph-grounded traversal produces provenance-backed causal narratives spanning infrastructure, orchestration, and governance layers, in contrast to the fragmented or unsupported explanations typical of log-only or LLM-only diagnostic approaches. The proposed architecture improves recommendation traceability, reduces ambiguity in AI-generated operational guidance, and provides a practical foundation for explainable, trustworthy AI-assisted operations in cloud-native DevOps platforms. Future work includes quantitative benchmarking against established AIOps datasets and integration with policy-as-code enforcement pipelines.