Retrieval-Augmented Generation versus Fine-tuning for Turkish Cultural Question Answering: A Comprehensive Evaluation and Analysis

This study presents the first comprehensive evaluation of Retrieval-Augmented Generation (RAG) systems versus fine-tuned language models for Turkish cultural question answering. While fine-tuned models achieve superior performance in short-context scenarios (F1=0.799), they fail to generalize effectively in long-document or cross-document settings. To address this, we introduce a RAG pipeline augmented with Turkish-specific adaptations, including morphology-aware chunking, cultural context weighting, and agglutination-aware answer normalization. These enhancements yield a 14.0\% improvement in overall RAG performance. We construct and evaluate a domain-specific Turkish cultural QA dataset and analyze the trade-offs between accuracy, scalability, and deployment feasibility. Our findings provide actionable guidelines for deploying QA systems in morphologically rich and culturally nuanced languages like Turkish.