A multimodal agent for ceramic management via LLM-driven knowledge and visual synergy

Ceramic heritage digitization has produced heterogeneous textual, visual, and structured data, while existing frameworks remain limited in structured extraction, fine-grained visual grounding, and evidence-based reasoning. Therefore, a multimodal ceramic analysis agent (CeramicAgent) is proposed, which integrates schema-constrained information extraction, high-quality visual segmentation based on the High-Quality Segment Anything Model (HQ-SAM), and graph-guided reasoning that connects textual, visual, and historical evidence. Experiments on a Palace Museum ceramic dataset show improved attribute extraction, reliable localization of micro-features, and coherent answers grounded in retrieved knowledge paths. The results indicate that the coordinated use of language, vision, and structured knowledge can support expert-oriented analysis of ancient ceramics, particularly when interpretation depends on degraded surface evidence, specialized terminology, and uncertain historical relations.