Prompt-to-Structure: Knowledge-Grounded LLM-Assisted Procedural Modeling of Qing Dynasty Architecture

Recent advances in large language models (LLMs) have opened new opportunities for integrating natural-language interaction with procedural content generation (PCG). However, heritage architectural modeling requires more than syntactically valid parameter generation, as generated forms must comply with historically grounded construction rules, component dependencies, and executable asset constraints. This study takes Qing-dynasty official timber architecture as a case study and presents Prompt-to-Structure , a knowledge-grounded framework that translates natural-language generation intent into minimal, executable procedural specifications. The framework integrates a reference corpus derived from architectural treatises and scholarship, a domain knowledge graph encoding architectural rules and applicability constraints, and a metadata-indexed 3D component library. Given a user prompt, retrieval-augmented grounding supports specification generation, while deterministic validation and component resolution ensure that the resulting parameters can be instantiated procedurally in Blender. Experiments show that, compared with unguided LLM baselines, the complete pipeline improves rule compliance while maintaining semantic alignment and reliable procedural executability. These results demonstrate the potential of constrained LLM–PCG workflows for rule-consistent heritage architectural generation and hypothesis visualization under historically grounded constraints.