Bridging Structured and Unstructured Knowledge: A Comparative Study of Retrieval-Augmented Generation and Open Knowledge Formats for Enterprise AI Systems

Enterprise artificial intelligence systems increasingly rely on large language models (LLMs) to answer questions, summarize records, and automate decision support across heterogeneous data estates that mix unstructured documents with structured, governed data assets. Two dominant strategies have emerged to ground LLM outputs in enterprise knowledge: Retrieval-Augmented Generation (RAG), which retrieves relevant unstructured passages from document corpora, and Open Knowledge Formats (OKF) — including the Resource Description Framework (RDF), the Web Ontology Language (OWL), JSON-LD, and knowledge graphs (KGs) — which encode facts as machine-interpretable, semantically typed triples. Existing literature typically evaluates these paradigms in isolation, leaving practitioners without a rigorous, multi-criteria comparison suited to enterprise constraints such as auditability, latency, and regulatory compliance. This paper presents a structured comparative analysis of RAG and OKF-based systems across ten enterprise-relevant dimensions: accuracy, hallucination rate, explainability, latency, scalability, cost, security, maintainability, enterprise readiness, and knowledge freshness. Building on this analysis, we propose a novel Hybrid Knowledge Routing Framework (HKRF) that classifies incoming queries by intent and structure, dynamically routes them to RDF/OWL-based knowledge graphs, document-oriented RAG pipelines, or both, merges heterogeneous evidence, and generates a final answer augmented with confidence scoring and source attribution. We describe the framework's architecture, routing algorithm, and mathematical formulation for confidence-weighted evidence fusion, and, going beyond a purely conceptual proposal, we implement a small, fully reproducible, offline pilot of the HKRF and evaluate it against RAG-only and OKF-only baselines on a hand-labeled 36-query benchmark spanning structured, unstructured, and hybrid queries. On this pilot, HKRF achieves higher overall answer accuracy than either single-paradigm baseline (mean token F1 of 0.646 versus 0.435 for OKF-only and 0.308 for RAG-only, McNemar's exact p = 0.0001 against both), with the largest relative gains on hybrid queries requiring both structured and unstructured evidence. The pilot substitutes a rule-based classifier, TF-IDF retrieval, and template-based answer synthesis for the trained classifier, dense retriever, and LLM generation stage of the full design, disclosed explicitly, because it was built and run in an offline sandbox; it validates the framework's routing and fusion mechanics on a small synthetic benchmark rather than constituting a large-scale, LLM-integrated, production evaluation, which remains future work. We discuss the framework's implications, its limitations, ethical considerations surrounding provenance and bias, and directions for future empirical work at greater scale. This study is positioned as a conceptual architecture and framework contribution, supported by an early empirical proof-of-concept, suited to applied and enterprise AI, knowledge engineering, and semantic technology venues