Retrieval augmented generation-driven information retrieval and question answering in construction management

Abstract

Construction management is a communication-intensive field, requiring prompt responses to queries from various stakeholders to ensure project continuity. However, retrieving accurate information from project documents is hampered by the mismatch in granularity between queries and vast contents and by inherent ambiguities in information. Large language models (LLMs) and retrieval-augmented generation (RAG) offer new opportunities to address the challenges. However, their effectiveness is limited by the segmentation of documents and insufficient consideration of engineers’ preferences. Therefore, we propose a novel paradigm: RAG for Construction Management (RAG4CM). It includes three components: 1) a pipeline that parses project documents into hierarchical structures to establish a knowledge pool; 2) novel RAG search algorithms; and 3) a user preference learning mechanism. The first two components enhance granularity alignment and RAG results by integrating document-level hierarchical features with raw contents. The preference learning realizes continuously improved responses along with user-system interactions. We developed a prototype system and conducted extensive experiments, demonstrating that the knowledge pool efficiently accommodates texts, tables, and images. RAG4CM realized a 0.924 Top-3 and 0.898 answer accuracy, surpassing both open-source frameworks and commercial products. In addition, preference learning further increases answer accuracy by 1.3 % to 9.5 %. Consequently, RAG4CM enables multi-source information retrieval in a user-friendly manner, improving communication efficiency and facilitating construction management activities.