Natural-Language to BIM: A Guard-Railed No-Code Pipeline for Reliable Model Creation

<p dir="ltr">Building Information Modeling (BIM) provides a single coordinated digital representation of the built asset and enables clash detection, 4D/5D planning, and data-driven decision making. However, practical BIM authoring remains expert-dependent: families, levels, types and parametrics make even small edits time-consuming and error-prone. Prior automation attempts (for eg. Dynamo, API add-ins, and unconstrained prompt-to-code) either demand programming skills or produce brittle, unreliable outputs. This paper presents a guard-railed, no-code pipeline that converts a single natural-language instruction in a simple text chat into a deterministic program script via a constrained LLM orchestrated on a server/website. The core idea of this paper is to use a reusable pre-prompt (contract) that fixes intent (goal), required fields, API commands, unit normalization and other necessary assumptions. This study discusses failure modes (vocabulary mismatch, geometric edge cases, prompt drift) and future intent to make the pipeline exhaustive for more complex structures as well.</p>