Abstract
The digital transformation of power dispatching and equipment operation increasingly relies on structured knowledge extracted from unstructured engineering documents. Existing extraction pipelines generally assume that source documents are architecturally complete and logically coherent. This assumption rarely holds in practice. Power engineering documents are often produced by reusing earlier files, which introduces missing chapters, omitted functional modules, and broken logical dependencies. Such defects propagate into extracted knowledge and degrade downstream applications. This paper proposes a topology-constrained method that combines Large Language Models with semantic topology graphs to evaluate the architectural integrity and content maturity of power engineering documents. Semantic fingerprinting and k-medoids clustering first identify homologous document families. A quality-verification stage then converts the cluster representative into an expert-checked reference architecture with mandatory, optional, and conditional functional units. The method maps prerequisite and data-production dependencies into a Directed Acyclic Graph while retaining feedback and interlock relations in a separate typed relation layer. A semantically gated partial graph-matching procedure identifies missing required nodes and dependencies without forcing incompatible nodes to match. A three-dimensional maturity model evaluates functional completeness, required-path preservation, and core element coverage. An evidence-constrained Graph Retrieval-Augmented Generation (GraphRAG) module further explains detected deficiencies and proposes corrections for engineering review. Experiments on 25,700 engineering documents from a provincial power grid enterprise report F1 scores of 0.84 to 0.87 and structural deficiency detection rates of 0.78 to 0.81 across two corpora. The framework provides an auditable quality gate for power knowledge engineering and a supervised mechanism for restoring document quality.