Brick-DICL: Dynamic In-Context Learning for Automated Brick Schema Classification

The article introduces Brick-DICL, a dynamic in-context learning framework designed for automated classification of the Brick schema in Building Management Systems (BMS). It features a two-stage architecture comprising metadata-RAG for enhancing LLM domain knowledge and class-RAG for narrowing classification options among 936 Brick classes, along with a multi-LLM filtering mechanism to improve prediction confidence. This approach significantly enhances classification accuracy and reduces manual verification effort, facilitating faster integration of standardized BMS across diverse datasets, which is crucial for interoperability in smart building technologies.