Title: Agentic Retrieval-Augmented Generation for Vietnamese Elementary History Education: A Hybrid Search, Cross-Encoder Reranking, and Pedagogical Assessment Framework
Authors: Hoang Ngoc Long, Nguyen Minh Duc, Trinh Thanh Do
Volume: 10
Issue: 9
Pages: 177-187
Publication Date: 2026/09/28
Abstract:
Providing accurate, engaging, and age-appropriate historical information to elementary students remains challenging because general-purpose large language models may generate unsupported facts, confuse historical periods, or use unsuitable language. This study presents Dai Viet Kids, a curriculum-grounded, agent-oriented retrieval-augmented generation framework for Vietnamese history education in Grades 4 and 5. The framework combines hierarchy-preserving breadcrumb chunking, sparse-dense hybrid retrieval with reciprocal rank fusion, cross-encoder reranking, intent routing, and persona-guided response generation with preliminary safety guardrails. We conducted a pilot evaluation using 46 curriculum-aligned questions and four adversarial robustness probes over 914 indexed textbook chunks. The full retrieval pipeline achieved Hit@1 of 74.00%, Hit@3 and Hit@5 of 94.00%, and a mean reciprocal rank at cutoff five (MRR@5) of 0.8233. BM25 alone achieved Hit@1 of 72.00%, Hit@5 of 94.00%, and an MRR@5 of 0.7827, so the difference between the two configurations is concentrated at Hit@3, with Hit@1 differing by a single query and Hit@5 identical; at this benchmark size no pairwise difference in Table 2 attains statistical significance, as Section 7 shows. Automated generation assessment yielded a faithfulness score of 85.05% and an answer-relevancy score of 87.94%. These findings provide preliminary evidence that hybrid retrieval and reranking can improve curriculum-grounded historical question answering. However, the small benchmark, limited robustness testing, and absence of teacher or learner evaluation mean that classroom effectiveness and child safety remain to be validated in future studies.