Scholarlink

<p>Large Language Models (LLMs) have demonstrated strong generative capabilities; however, their direct application in educational tutoring systems is limited by hallucination, lack of curriculum alignment, and unreliable factual grounding. Retrieval-Augmented Generation (RAG) has emerged as a promising solution, yet most existing RAG frameworks are designed for open-domain knowledge and are not optimized for curriculum-constrained educational settings. In this work, we propose a domain-adapted RAG-based intelligent tutoring system that treats NCERT textbooks as authoritative knowledge sources. The proposed pipeline integrates semantically informed overlapping chunking, dense embedding-based retrieval, and a precision-oriented semantic reranking mechanism to mitigate retrieval noise prior to generation. Unlike generic RAG implementations, our approach explicitly addresses educational constraints by prioritizing factual correctness, conceptual coherence, and syllabus alignment. We evaluate the system using retrieval-focused metrics (Precision@K, Recall@K) and generation-quality measures (BLEU, ROUGE-L, and BERTScore), demonstrating consistent improvements over retrieval-only and generation-only baselines. The results indicate that semantic reranking significantly enhances contextual relevance and reduces hallucinated responses while maintaining low response latency suitable for real-time tutoring. This study establishes a reproducible and scalable framework for grounding LLM-based tutoring systems in curriculum-aligned educational resources, offering a practical pathway toward reliable and explainable AI-assisted learning environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Evaluating Large Language Models in Retrieval-Augmented Tutoring Systems: Methods and Emerging Tools

  • 期刊论文

  • 发表日期：

    2026年3月11日

查看全文
  • Owais Bhat,
  • Zubair Jeelani,
  • Syed Tanzeel Rabani,
  • Syed Mohsin Saif,
  • Nawaz A. Lone

摘要

Large Language Models (LLMs) have demonstrated strong generative capabilities; however, their direct application in educational tutoring systems is limited by hallucination, lack of curriculum alignment, and unreliable factual grounding. Retrieval-Augmented Generation (RAG) has emerged as a promising solution, yet most existing RAG frameworks are designed for open-domain knowledge and are not optimized for curriculum-constrained educational settings. In this work, we propose a domain-adapted RAG-based intelligent tutoring system that treats NCERT textbooks as authoritative knowledge sources. The proposed pipeline integrates semantically informed overlapping chunking, dense embedding-based retrieval, and a precision-oriented semantic reranking mechanism to mitigate retrieval noise prior to generation. Unlike generic RAG implementations, our approach explicitly addresses educational constraints by prioritizing factual correctness, conceptual coherence, and syllabus alignment. We evaluate the system using retrieval-focused metrics (Precision@K, Recall@K) and generation-quality measures (BLEU, ROUGE-L, and BERTScore), demonstrating consistent improvements over retrieval-only and generation-only baselines. The results indicate that semantic reranking significantly enhances contextual relevance and reduces hallucinated responses while maintaining low response latency suitable for real-time tutoring. This study establishes a reproducible and scalable framework for grounding LLM-based tutoring systems in curriculum-aligned educational resources, offering a practical pathway toward reliable and explainable AI-assisted learning environments.