The expansion of descriptive and essay-type assessment in South Korea’s 2022 Revised National Curriculum and 2028 College Admissions Reform Plan has intensified the need for reliable scoring by Korean Language Arts teachers. Existing rater training approaches face a persistent dual limitation: face-to-face workshops are constrained by spatiotemporal barriers and their one-off nature, while computer-based systems provide only superficial feedback that fails to address the cognitive processes underlying scoring decisions. This study proposes Retrieval-Augmented Generation (RAG) as a technology capable of bridging this gap and presents the design of a RAG-based rater training program. Drawing on Cognitive Apprenticeship, Frame-of-Reference training, and Deliberate Practice theory, six core design principles were derived through a two-axis analysis integrating theoretical mechanisms with design limitations identified in prior programs. These principles were translated into a seven-stage program following a 2-Round learning path (pretest–intervention–posttest). The program’s RAG system, built on open-source Korean-language LLMs (EXAONE 3.5 7.8B and SOLAR 10.7B) deployed on institutional on-premise servers, differentiates the roles of retrieval and generation across stages: RAG ensures knowledge consistency through precise case retrieval from an expert scoring database, while the LLM performs complex reasoning and personalized feedback synthesis. An adaptive feedback mechanism adjusts content and tone based on diagnosed scoring tendencies (leniency or severity), and a structured What-Where-Why scoring rationale requirement in the second scoring round promotes metacognitive reflection by making the cognitive process of scoring—not merely its numerical outcome—a target of training and feedback. As a design-stage study, this work presents the program’s architecture in implementable detail; empirical effectiveness verification through field deployment with practicing teachers is reserved for subsequent research.