While large language model (LLM)-based QA systems are prevalent, there remains a significant gap in research on effectively integrating multimodal inputs with retrieval-augmented generation (RAG). Questions on space research websites range from administrative to complex technical queries, yet baseline LLMs often perform poorly due to their limited parametric memory and pre-training data. This paper proposes MERIT, which enhances RAG by processing both text and image inputs to build a comprehensive multimodal knowledge base, thereby significantly improving retrieval capabilities. Experiments comparing our model’s answers with human-validated ground truths show substantial improvements over state-of-the-art LLM baselines in ROUGE and BERT scores. These results underscore MERIT’s potential to address diverse QA scenarios beyond conventional text-only RAG systems.