No Pre-Trained Knowledge, No Hallucination: Answering Student Questions with Sentences Retrieved from Teacher-Provided Documents

Hallucination-the tendency of large language models (LLMs) to generate incorrect or fabricated information-poses a significant risk to student learning. Although retrieval augmented generation (RAG) has been proposed to mitigate hallucination by grounding LLM outputs in external sources, mitigation does not equate to elimination. Existing RAG implementations still let the LLM draw on its pre-trained knowledge when composing an answer, so hallucinated content can persist even when relevant documents are retrieved.

This paper introduces RAG-S , a sentence-level retrieval framework that removes this risk at its source. Unlike conventional RAG, which segments documents into fixed-size chunks with arbitrary overlap, RAG-S indexes text at the sentence level, treating each sentence as a selfcontained retrieval unit. During retrieval, RAG-S compares a student's question against every indexed sentence for semantic similarity, returning those that exceed a predefined threshold in descending order of relevance.

An optional contextual expansion mechanism retrieves adjacent sentences to preserve coherence. Because these retrieved sentences are the original, unaltered text, they are already a correct, hallucination-free answer that both teachers and students can read directly. Generation is therefore optional: when it is used, the LLM is restricted to these retrieved sentences and does not draw on its pre-trained knowledge, and RAG-S shows the source sentences before the generated answer so students can verify and interpret the information themselves.

This design reinforces a fundamental principle of traditional education: engaging with and critically evaluating original source material. It also leverages AI for precise, efficient, and exhaustive sentence-level retrieval. By anchoring every response to teacher-provided text rather than pre-trained knowledge, RAG-S removes hallucination rather than merely reducing it.

Experimental results show that RAG-S consistently retrieves relevant sentences and generates source-faithful responses across multiple LLMs. The framework offers a principled approach to integrating AI into education: it improves factual accuracy, promotes critical thinking, and supports rather than replaces the learning process.