CONSTRUCTA: Automating Commercial Construction Schedules in Fabrication Facilities with Large Language Models

TL;DR

CONSTRUCTA tackles the challenge of automating construction schedules in complex projects by integrating three core components: Static Retrieval-Augmented Generation for domain knowledge, Contextual Knowledge RAG for dynamic, context-sensitive input, and Construction RLHF with CPA-DPO to align outputs with expert preferences. On a proprietary dataset of 4,340 semiconductor activities, it achieves substantial improvements in missing value prediction, dependency analysis, and automated planning, outperforming baselines and highlighting strong discipline-, level-, and area-specific gains. The approach demonstrates how domain-informed retrieval, hierarchical context sampling, and human-guided optimization can yield robust, scalable, and adaptable scheduling for construction workflows, with clear path toward integration with industry tools. These results suggest significant practical impact for automated planning in large-scale construction environments and inform future research on domain-specific LLM customization and deployment strategies.

Abstract

Automating planning with LLMs presents transformative opportunities for traditional industries, yet remains underexplored. In commercial construction, the complexity of automated scheduling often requires manual intervention to ensure precision. We propose CONSTRUCTA, a novel framework leveraging LLMs to optimize construction schedules in complex projects like semiconductor fabrication. CONSTRUCTA addresses key challenges by: (1) integrating construction-specific knowledge through static RAG; (2) employing context-sampling techniques inspired by architectural expertise to provide relevant input; and (3) deploying Construction DPO to align schedules with expert preferences using RLHF. Experiments on proprietary data demonstrate performance improvements of +42.3% in missing value prediction, +79.1% in dependency analysis, and +28.9% in automated planning compared to baseline methods, showcasing its potential to revolutionize construction workflows and inspire domain-specific LLM advancements.

Figures (9)

• Figure 1: Overview of the Constructa system. (a) The initial construction schedule is created by experts and refined with contextual activity and site samples. (b) Contextualized activity aggregates hierarchical, first-order, and sequential relations. (c) Knowledge vectorization embeds and retrieves construction knowledge for optimization. (d) Construction preference alignment uses RLHF to align schedules with expert rules and preferences.
• Figure 2: Illustration of the CPA-RLHF process. ① Raw contexts and rules are input for comprehension. ② The Plan Agent refines these into filtered contexts and rules. ③ Completions are evaluated and stored in the Preference Database. ④ The Expert Agent aligns outputs with project preferences. Part (a) collects data for preference model training, and part (b) aligns preferences for accurate planning.
• Figure 3: Performance Comparison Across Levels and Areas. This plot shows the performance of various metrics, including Basic Context, Static RAG, Knowledge RAG, and Construction RLHF, for three tasks (Automated Planning, Dependency Analysis, and Missing Value Prediction) across different levels and areas.
• Figure 4: Context length distributions for AP, DA, and MVP sources, highlighting reductions achieved through CPA-DPO. The shorter contexts effectively maintain performance while improving efficiency in schedule optimization.
• Figure 5: Distribution of degree and maximal hop for dependency graph nodes. The left plot shows the degree distribution, reflecting task interconnectivity, while the right plot presents the maximal hop distribution, highlighting long-range task dependencies.