Effective presentation slide creation is crucial for impactful communication, yet fully automating this task with AI is insufficient. Hybrid human-AI solutions often perform worse than pure AI or human creation due to overreliance on AI. To address this, we develop design principles for configuring human-AI hybrid systems in complex knowledge tasks using a design science research approach. Our prototype, NarrativeNet Weaver, leverages an underutilized corpus of existing presentation slides, applying generative AI advances in hybrid dense embedding and graph-based retrieval techniques. Evaluated through 15 think-aloud sessions and 73 user trials, users with NarrativeNet Weaver exhibit greater engagement and achieve equal or improved slide quality compared to those using a ChatGPT-based chatbot with a vector database. We contribute design knowledge for human-AI systems for complex multimodal content and offer a new approach to retrieving and visualizing existing slides, enhancing the utilization of valuable but underused resources.