Graph-Oriented Artificial Intelligence: An Integrated GNN–RAG–KG–LLM Architecture for Developing Trustworthy Enterprise Artificial Intelligence

Abstract

This study aimed to design and empirically evaluate an integrated graph-oriented architecture combining Graph Neural Networks (GNNs), Retrieval-Augmented Generation (RAG), Knowledge Graphs (KGs), and Large Language Models (LLMs) to improve the trustworthiness of enterprise artificial intelligence. This applied, developmental, and quantitative evaluation study was conducted in Tehran, Iran, in 2026 with 48 experts selected purposively from technology companies, universities, financial institutions, telecommunications organizations, and enterprise information technology departments. The proposed architecture integrated an enterprise knowledge graph, GNN-based relational representation learning, RAG-based evidence retrieval, and LLM-based natural-language generation. Its performance was compared with LLM-only, LLM + RAG, and LLM + RAG + KG configurations. Data were collected through standardized benchmark queries, automated computational metrics, provenance logs, and a researcher-developed expert questionnaire. Repeated-measures analysis of variance, pairwise comparisons, effect-size estimation, correlation analysis, and ablation analysis were used. Significant differences were observed among the four architectures for answer correctness, F = 126.41, p < 0.001, partial η² = 0.729; context relevance, F = 94.83, p < 0.001, partial η² = 0.669; faithfulness, F = 139.27, p < 0.001, partial η² = 0.748; groundedness, F = 153.62, p < 0.001, partial η² = 0.766; evidence traceability, F = 241.53, p < 0.001, partial η² = 0.837; hallucination rate, F = 172.84, p < 0.001, partial η² = 0.786; and expert-rated trustworthiness, F = 118.76, p < 0.001, partial η² = 0.717. The integrated architecture significantly outperformed all reduced configurations. Expert trustworthiness was positively correlated with groundedness (r = 0.78), faithfulness (r = 0.74), evidence traceability (r = 0.81), and correctness (r = 0.72), and negatively correlated with hallucination rate (r = −0.76), all p < 0.001. Integrating GNN, RAG, KG, and LLM components produced a more accurate, grounded, traceable, and trustworthy enterprise AI system while substantially reducing hallucination, supporting the value of graph-oriented architectures for reliable organizational AI applications.