Running summarizing tools on a new article

Abstract: Abstract School students increasingly use large language models as study aids, yet fluent but unfounded answers pose a distinctive risk in education, because a learner who asks a question is not positioned to verify the reply. Retrieval-augmented generation (RAG) mitigates this hallucination risk but does not eliminate it, and the evidence base comes almost entirely from high-resource languages. This study examines hallucination in textbook-grounded question answering for Uzbek, an agglutinative and under-resourced language. In a 30,628-word official history textbook, 55.5% of the 7,867 distinct word forms occur exactly once, so query and document surface forms frequently fail to match. We construct UzHistQA (Uzbek History Question Answering), an openly available dataset of 56 question–answer pairs in which each answer is linked to a gold passage and approximately one fifth of questions are deliberately unanswerable. Four configurations are compared: an ungrounded model, dense-retrieval RAG, hybrid retrieval with reranking, and hybrid retrieval with enforced citation and abstention. Grounding reduced unsupported assertions from 91% of responses to 9% (p < 0.001) and fabricated answers to unanswerable questions from eleven of twelve cases to one; enforced abstention eliminated fabrication entirely, at a coverage cost of 14%. Hybrid retrieval improved recall yet did not reduce hallucination, indicating that retrieval-stage gains need not propagate to the generated answer. The study contributes the first open evaluation resource for grounded question answering in Uzbek and evidence that retrieval quality and faithfulness must be measured separately.