Game-Specific RAG-LLM for Automated QA Generation and Player Support

• Gaeun Park,
• Zhongjie Jin,
• Yejin Kim,
• Beomjoo Seo,
• Shinjin Kang,

Vol. 20, No. 3, March 31, 2026
10.3837/tiis.2026.03.002, Download Paper (Free):

• Retrieval Augmented Generation
• Large Language Model
• Quality Assurance
• Vector database
• Knowledge Graph Database

Abstract

The dramatic increase in complexity across both the systemic and scale aspects of modern digital games poses a severe challenge to traditional manual Quality Assurance (QA) process, which are inherently inefficient and susceptible to human error. Addressing this, we propose a Retrieval Augmented Generation (RAG)-Large Language Model (LLM) system specialized for complex game domains. This model is designed with two core objectives: to automate the generation of QA test cases for improved development efficiency and to provide a Q&A chatbot for game players to improve information accessibility. We constructed a comprehensive knowledge base using data from the MMORPG World of Warcraft. Based on this dataset, we designed and comparatively analyzed three distinct RAG architectures operating under an LLM framework: (1) a baseline RAG utilizing metadata-based filtering, (2) a semantic chunking and search-based RAG, and (3) a structured Knowledge Graph (KG)-based RAG. Performance evaluation conclusively demonstrated that the KG-based RAG architecture achieved superior results in terms of both answer accuracy and response speed. Furthermore, a user survey involving actual game players yielded positive feedback regarding the practicality and satisfaction of its Q&A functionality. This research thus confirms that our proposed KG-based model can substantially contribute to innovating the game QA pipeline and significantly enhancing user experience, promising a positive impact on overall game development and service delivery.

Statistics

Cite this article