AI-Enhanced Paragraph Retrieval in Academic Databases: Empirical Evaluation of Stability, Robustness, and Adaptive Optimization for Smart Learning Environments

The performance of an AI-enhanced paragraph retrieval system in academic databases is evaluated, focusing on stability, robustness, and adaptive optimization for smart learning environments. Empirical experiments across diverse retrieval conditions analyze system behavior under time-series stability tests, input variability, and multi-source constraints. Results demonstrate consistent performance across time slots and account types (e.g., personal vs. institutional users), though sensitivity to minor input variations (e.g., synonym variations and typographical errors) and high-load bottlenecks (e.g., “no data” errors, timeouts) is observed. Key metrics-including precision (0.40-0.45), recall (0.22-0.52), F1-score (0.20-0.45), and NDCG (0.93-0.96)-highlight trade-offs between retrieval modes: “Paragraph Theme” prioritizes high recall with broader coverage, “Subtitles Only” balances precision and recall, and “Title & Subtitles” offers broad coverage at the cost of lower accuracy. Input variations cause significant fluctuations in retrieval volume, underscoring limitations in keyworddependent retrieval. Hybrid retrieval strategies combining sparse/dense embeddings, distributed computing architectures, and explainable AI (XAI) techniques are proposed to enhance robustness and transparency. Findings inform the design of adaptive, user-centric retrieval systems for educational technologies, emphasizing semantic robustness and scalable infrastructure in AI-driven academic platforms.