<![CDATA[The integration of Large Language Models into game engines is hindered by cloud dependency and high latency. This study proposes a fully localized Retrieval-Augmented Generation (RAG) framework using the Llama 3.2 Small Language Model to generate role-playing game quests and maintain character personalities without parameter fine-tuning. Operating within a C++ environment under strict hardware constraints (primarily CPU-bound), the methodology evaluates three retrieval methods (MiniLM, FastText, and BPE Tokenizer) combined with a memory-efficient JSON Vector Database. System effectiveness was measured using BLEU, ROUGE, and user evaluations. Results show the initial BPE Tokenizer achieved the lowest quest generation time of 152.01 seconds, the fastest average response time of 23.53 seconds, the highest BLEU score of 0.0118, and a peak persona consistency score of 3.70. However, the relatively long response times remain a primary weakness hindering real-time immersion. Furthermore, a critical "Stopping Condition Dilemma" emerged; lacking engine awareness, the model failed to detect narrative conclusions. This caused generation loops that spiked processing times up to 220.59 seconds. Future research must integrate strict, state-based logic triggers from the game engine to prevent context collapse and optimize inference to reduce latency.]]>
Copyrights © 2026
