Graph retrieval augmented generation for multimodal construction and demolition waste data integration toward traceable disposal recommendations

Construction and demolition waste (C&DW) management is central to urban renewal and sustainable construction, yet current intelligent solutions often suffer from fragmented multimodal data, weak knowledge linkage, and limited policy interpretability. Image recognition models can classify C&DW materials but cannot integrate regulatory clauses, disposal processes for explainable reasoning, while text centered large language models (LLM) retrieval lacks visual inputs and knowledge graph based path reasoning. This paper proposes a multimodal framework that integrates Contrastive Language-Image Pre-training (CLIP) ViT-B-32 for zero-shot visual evidence extraction, a lightweight Neo4j based C&DW knowledge graph for structured domain modeling, a hybrid Graph Retrieval Augmented Generation (GraphRAG) workflow for joint subgraph and vector retrieval, and GPT-4o for evidence-anchored recommendation generation. The framework links on-site images with regulatory clauses and disposal pathways to generate disposal recommendations, compliance alerts, and clause-level citations. Experimental results show that the proposed framework achieves 71% end-to-end accuracy, 100% citation coverage, and a 0% hallucination rate, outperforming single modality baselines, while remaining scalable and easy to deploy.