Benchmarking retrieval strategies for curriculum-aligned retrieval-augmented generation using NCERT-RAG-Eval STEM textbooks

LLM-based tutoring systems risk producing wrong or hallucinated explanations unless grounded in trusted educational material. We propose NCERT-RAG-Eval, a curriculum-aligned benchmark for evaluating retrieval-augmented generation (RAG) against official NCERT textbooks, the most widely used books in Indian STEM education. The framework indexes a 237-chunk NCERT corpus spanning five chapters and compares lexical (BM25), dense semantic (FAISS, ), and hybrid (reciprocal rank fusion) retrieval on recall, precision, and faithfulness. On 88 gold-annotated questions spanning motion, light, electricity, and magnetic effects, hybrid retrieval performs best (Recall 0.818, Precision 0.348), closely followed by FAISS (0.807, 0.341) and clearly ahead of BM25 (0.682, 0.280); on a paraphrased subset, Hybrid shows no Recall drop at all, while FAISS drops the most. Growing the corpus with non-duplicated distractor content leaves Precision essentially flat even at more than four times the original corpus size, unlike the sharp degradation seen under duplication-based growth. A small generation-level pilot with an open-weight LLM (Llama-3.1-8B), run on the original 40-question Motion-only benchmark, found closed-book generation more accurate (94.9%) than RAG-grounded generation (82.5%) – a difference not statistically significant at this sample size – with RAG errors traced to fragmented chunk boundaries, a text-extraction artifact, a misleading retrieval, and an arithmetic slip, rather than to retrieval being unhelpful in general. The benchmark, questions, and pipeline are publicly released to support reproducible research on trustworthy AI-supported learning.