Related Experiment Video
Updated: Sep 2, 2026
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
Optimising Large Language Models-Based Automated Question Generation for Dental Pharmacology: A Quantitative
Fumihiko Yoshino1, Ayaka Yoshida2, Shun-Suke Takahashi1
1Department of Pharmacology, Kanagawa Dental University, Yokosuka, Japan.
Introduction:
The application of Large Language Models (LLMs) for automated question generation in dental education is hindered by inefficient data management, where conventional systems using a single data file incur high Application Programming Interface (API) costs and slow performance due to excessive token consumption. This study aimed to quantitatively validate the performance benefits of an optimized, multi-file JavaScript Object Notation (JSON)-based architecture over this conventional approach.
Materials And Methods:
We developed a system using Gemini 2.5 Pro and compared two architectures: a conventional unmodified method (single Markdown file) and our proposed modified method (multiple, purpose-specific JSON files). We measured API processing time and token counts for generating questions (n = 5 trials per condition) and the data were analysed using a two-way analysis of variance (ANOVA).
Results:
The modified JSON-based architecture demonstrated superior efficiency. For a 10-question task, it significantly reduced API processing time (97.0 vs. 115.6 s, p < 0.001) and reduced input token consumption by approximately 88% (95 197 vs. 813 702 tokens, p < 0.001) compared with the unmodified method. All generated question sets passed automated structural and formatting validation.
Conclusion:
In conclusion, a structured, multi-file JSON architecture dramatically enhances the efficiency and scalability of LLM-based educational tools. This approach provides a practical framework for applying AI in dynamic fields such as pharmacology and may help reduce operational burden in modern curriculum development.
