Accessing Building Information Model (BIM) data encoded in Industry Foundation Classes (IFC) remains difficult for non-expert stakeholders due to heterogeneous property sets and nested relationships. This paper presents a constrained text-to-graph querying framework that transforms IFC data into a labeled property graph, queried via Cypher, while project-specific Information Delivery Specification (IDS) requirements define admissible entity labels, relationships, and property keys. A regular-expression grammar restricts Large Language Model (LLM) generation to executable, schema-compliant queries. Evaluation shows grammar constraints eliminate schema hallucinations, achieving 100% Semantic Compliance and improved execution accuracy, with competitive performance from smaller open-weights models.
Information Extraction (IE) from Architecture, Engineering, and Construction (AEC) drawings is hindered by inefficient manual processes. Although layout detection is a key prerequisite for downstream semantic interpretation—organizing graphical and textual hierarchies—it remains under-explored in this domain. Most general document layout analysis models are optimized for text-centric documents. Consequently, their performance on complex engineering drawings remains unvalidated.
To address this gap, we construct a custom dataset and benchmark five diverse deep learning architectures on AEC-specific layouts. Results demonstrate that RF-DETR achieves state-of-the-art performance, with an mAP@50 of 0.949, proving the DETR architecture's superiority in capturing high-density, sparse graphical layouts. Additionally, Vision-Language Models like Qwen3-VL demonstrate strong interpretive ability, achieving an F1-score of 0.911. In contrast, models pre-trained on general document datasets exhibit domain interference, leading to degraded performance. This research confirms that distinct data features prevent effective model transfer from general documents, establishing a robust technical foundation for automated IE.
Vision–language models offer a promising alternative to supervised computer vision by enabling open-set object recognition without retraining or large annotated datasets. Recent advances in open-vocabulary object detection models, such as GroundingDINO, support object grounding through natural language prompts; however, their sensitivity to prompt formulation remains underexplored in the AEC domain. This study evaluates the reliability of GroundingDINO for construction objects detection by systematically analysing prompt variation. Lexical variation and prompt multiplicity are examined. A custom dataset of 48 construction objects was evaluated under eight prompt configurations. Results show that lexical variation significantly affects detection performance, with “regular base term” achieving the highest F1 score of 56.1%, and multiple prompting decreases performance by about 2%, highlighting the effectiveness and limitations of AEC applications.
Routine construction site Health and Safety (H&S) inspections remain manual, experience-driven, and weakly digitised. While Computer Vision (CV) has advanced construction safety, particularly for real-time monitoring, it has rarely been investigated to support everyday inspection practices. This paper addresses this gap by analysing a routine inspection process and formalising current practice using Business Process Model and Notation (BPMN) 2.0 to make roles, decisions, and information exchanges explicit and expose key limitations. The study then proposes a CV-enabled inspection process that structures the collection and interpretation of visual evidence and suggests compliance-related observations to the H&S Manager, while preserving their assessment responsibility.
This paper presents a framework for information retrieval from the heterogeneous data sources in the construction sector. The core of the approach is a Large Language Model (LLM)-driven multi-agent system (MAS) consisting of specialized sub-agents and tools, each assigned to a specific data type. A host agent orchestrates the overall process. As a proof of concept, an implementation for Building Information Modeling (BIM) models and technical standards is presented. Challenges arise at the interfaces between LLM agents and the retrieval mechanisms. These are addressed by employing LangChain as an agent framework and the usage of the Model Context Protocol (MCP) as a novel standardized interface. The results demonstrate that such a MAS can answer complex queries that require information from multiple heterogeneous sources by combining several retrieval methods. The study provides a foundation for developing increasingly comprehensive systems through the application of various adaptation and optimization mechanisms.
This paper proposes a three-stage workflow to prioritise reusable building product groups in Dutch demolition outflows, addressing the gap between impact potential and implementation feasibility. Stage 1 screens 139 groups using an impact proxy combining mass outflow and the environmental cost indicator, shortlisting ten candidates. Stage 2 applies a five-criteria multi-criteria analysis with weights from eight experts using the Best–Worst Method, where supply–demand balance dominates (0.296) and sustainability ranks lowest (0.137). Stage 3 links ranking to barrier diagnosis through a masonry case, identifying contamination, certification, and coordination challenges. The workflow supports identifying high priority product groups for early-stage reuse planning.
Inner-city construction projects face severe spatial constraints during the finishing stage, often leading to mixed construction and demolition waste (C&DW) handling and reduced circular performance. This paper proposes a project-level decision-support framework to evaluate outbound C&DW logistics strategies, addressing the lack of project-level comparative tools. The framework integrates economic, environmental, and social indicators in a scenario-based comparison of conventional handling and CCC-enabled reverse logistics. Applied to three Dutch projects, results show cost and emission reductions, with slightly higher logistics intensity. The framework supports ex ante decision-making under spatial constraints.
Information management of construction site documentation is crucial for cost control in construction companies. Currently, sites are often managed with traditional tools, leading to slow archiving processes and data inconsistencies. This highlights the need for a more efficient document-management system, where the IFC format can act as a standardized structure for organizing the data recorded in the documents. This paper proposes a framework to optimize the use of site documentation by mapping and classifying key information. The structured extraction of data, as man-hours, enables integration with management control systems and BIM models, supporting more accurate evaluation of construction cost performance.
Circular built environment relies on reliable, reusable information across life-cycle stages; however, current initiatives remain fragmented. Identifying interoperability as a primary barrier, this paper draws on insights from a multi-stakeholder workshop to present a systemic roadmap. The research consolidates expert perspectives into sequenced, mutually dependent interventions across five domains: shared strategic vision; aligned standards and requirements; organisational and human capability; incentives and value realisation; and integrated system architecture. By clarifying cross-domain dependencies, this framework provides an actionable basis for prioritising pilots, policy mechanisms, and digital infrastructure to scale circular practices beyond isolated demonstrators.
Driven by LLM advances, Question Answering (QA) systems for Building Information Models (BIM) have proliferated, yet no standardized methods exist for cross-study comparison. We present an automated evaluation pipeline employing LLM judges to assess BIM-QA answers across five quality criteria and validate it through inter-rater agreement analysis with two domain experts and two LLM judges on 74 IFC-Bench question-answer pairs. LLM judges achieve high inter-rater reliability (Krippendorff's α = 0.70--1.00), exceeding human experts (α = 0.32--0.57), and show good agreement with expert consensus (α = 0.48--0.85), positioning the pipeline as a suitable solution for automated, reproducible BIM-QA benchmarking.
