Abstract
Retrieval-augmented generation enables precise educational question-answering by combining retrieval with natural language generation. Limited evaluation exists for educational RAG systems, particularly for curriculum-based applications in diverse languages. We introduce systematic evaluation using Turkish Ministry of Education textbooks, comparing embedding architectures, generation models, and optimization techniques across educational questions. Results establish multilingual-e5-large-instruct and BAAI/bge-m3 as optimal embedding models, demonstrate Qwen3 and Turkish-Gemma-9b's superior generation performance, and show ensemble methods yield 3.3% improvements. This provides evidence-based recommendations for educational RAG development across different languages and curricula.