Existing Retrieval-Augmented Generation (RAG) systems face critical limitations when processing scientific literature containing multi-modal content. Traditional document processing separates figure references from their corresponding images and data visualizations, breaking semantic relationships essential for a comprehensive understanding. In this paper, we propose an Agentic RAG framework that preserves and exploits cross-modal and cross-source semantic connections through systematic figure-text linking and dynamic modal weighting. The proposed approach builds associative relationships between textual references and visual content during pre-processing, enabling unified retrieval across different information modalities. Experimental validation demonstrates significant improvements in query performance with particularly strong results for hybrid questions requiring integrated understanding of text, visual, and structured data.