Comparative Analysis of Chunking Strategies for Turkish Multimodal RAG Systems
2026 34th Signal Processing and Communications Applications Conference (SIU), İstanbul, Türkiye, 7 - 10 Temmuz 2026, ss.1-4, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/siu71813.2026.11636565
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Sayfa Sayıları: ss.1-4
- Sivas Cumhuriyet Üniversitesi Adresli: Evet
Özet
Traditional Retrieval Augmented Generation (RAG) systems are often insufficient in preserving information integrity when dealing with complex documents that contain tables, graphs, and irregular layouts. To overcome these structural problems, this study provides a comparative analysis of four different chunking strategies: Naive, Dynamic, Vision, and Hybrid RAG. The Hybrid architecture is based on a multi-stage workflow integrating structure—aware intelligent text processing-which preserves the table hierarchy—with targeted visual vector generation. The performance of these strategies was evaluated on a dataset comprising 46 corporate integrated reports and 1,380 question-answer pairs, using an LLM-as-a-Judge approach across Context Recall, Precision, and Accuracy metrics. The results demonstrate that Hybrid RAG effectively overcomes the limitations of unimodal systems. Compared to the best unimodal baseline, it achieved 0.82 Context Recall, 0.82 Precision, and 0.71 Answer Accuracy.