PDF 논문 RAG, 텍스트만으로 충분할까? - Gemini embedding 002 임베딩 검색 실험

This research challenges the conventional text-only approach to academic paper RAG by demonstrating that image-based indexing significantly outperforms text embeddings alone. Using Gemini embedding-2-preview, experiments reveal that approximately 36% of visual information from figures and diagrams is lost in text-only extraction, with image indexes achieving 0.719 MRR versus 0.631 for text. Organizations handling visually-rich documents should reconsider default RAG pipelines that rely solely on text extraction and adopt multimodal indexing strategies to preserve critical contextual information.