Current smart indoor environment management face challenges in integrating heterogeneous data and supporting interactive, occupant-centric control. Existing methods lack semantic interoperability and struggle to capture occupant preferences dynamically. This chapter presents an AI agent framework that leverages semantic modelling to unify multi-source building data and integrates a vector-graph-based Retrieval-Augmented Generation (V-G RAG) engine for querying domain knowledge. Large Language Models (LLMs) facilitate natural language interaction and preference interpretation. Experimental deployment in an operational office building demonstrates a significant improvement in query accuracy over baseline methods. The proposed framework enables scalable, interpretable, and adaptive indoor environment optimisation, advancing occupant-centric control in smart indoor environment management.