Augmenting LLM reasoning with semantic web for knowledge sharing in off-site construction

Effective knowledge sharing and reuse are critical in off-site construction (OSC). However, their adoption remains limited due to the unstructured, fragmented, and project-specific nature of OSC knowledge. Although the recent introduction of large language models (LLMs) with retrieval-augmented generation (RAG) strategies has improved information accessibility, these models are still constrained by their static databases and lack sufficient reasoning capability to handle dynamic product configuration in OSC projects. Therefore, this paper proposes an adaptable semantic web-integrated RAG (SWiRAG) framework that combines semantic web technologies (SWTs) with LLMs to improve knowledge compatibility in OSC. In this framework, an ontology model and reasoning engine are embedded in the RAG pipeline. The ontology model can represent explicit knowledge and project-specific facts, while the reasoning engine can dynamically enrich them with implicit relationships and attributes for logical inference over static database. Furthermore, an optimal retrieval algorithm is developed to enable flexible, context-aware retrieval. Unstructured datasets such as illustrative drawings are also linked to the ontology to provide supplementary information to users. To evaluate the key components of SWiRAG, a prototype system was developed and tested on inter-modular connection designs using standard metrics of faithfulness, relevance, and recall across multiple LLMs. The results show that the proposed dynamic method significantly outperforms static RAG approaches, with average improvements of 38% in faithfulness, 31% in relevance, and 120% in recall. Based on the extended IFC datasets, SWiRAG also has better performance compared to existing graph-based retrieval frameworks. Furthermore, 26 participants from the construction domain evaluated the system, providing an average usability score of 82.8, which reflects excellent performance relative to standard benchmarks. This study demonstrates that ontology-based reasoning can effectively augment LLMs, enhancing the accessibility, reusability, and scalability of knowledge in OSC and providing a foundation for improved intelligent and adaptive construction knowledge systems.