Knowledge-Enhanced Large Language Models for Domain-Specific Intelligent Applications

Abstract

Large language models have become general-purpose computational infrastructures, yet their deployment in high-stakes domains requires more than scale and fluency. Domain-specific intelligent applications must operate with reliable factual grounding, traceable reasoning, institutional policy compliance, and robust behavior under distributional shift. This paper examines knowledge-enhanced large language models as socio-technical systems rather than isolated model artifacts. It argues that domain specificity emerges through the articulation of model weights, external knowledge substrates, retrieval mechanisms, orchestration layers, human review, governance processes, and evaluation infrastructures. The paper first reviews conceptual foundations, contrasting parametric memory with retrieval-augmented and knowledge-graph-based approaches. It then analyzes architectural trade-offs involving latency, cost, updateability, transparency, and failure containment. Subsequent sections discuss knowledge acquisition and curation, tool-augmented reasoning, deployment and sustainability constraints, governance and fairness obligations, and cross-sector applications in medicine, law, engineering, and public administration. A central claim is that knowledge enhancement is not merely an accuracy intervention but an institutional design choice that redistributes authority among model providers, domain experts, data stewards, auditors, and end users. The paper also considers evaluation as an assurance practice, emphasizing adversarial testing, provenance, uncertainty communication, and continuous monitoring. It concludes that sustainable knowledge-enhanced systems require modular architectures, clear accountability pathways, and governance mechanisms that treat domain knowledge as a living infrastructure rather than a static training resource.