Go to page

Bibliographic Metadata

Links
Download
Classification
Abstract
  • The reuse and adaptation of SysML modules and model fragments is central to efficient model-based systems engineering, yet it doesn't remain easy in practice due to heterogeneous documentation, fragmented repositories, and the high effort required to produce syntactically valid and tool-agnostic models. Large language (LLM) models offer automation potential, but unconstrained generation produces hallucinations and syntactically invalid SysML v2 that cannot be parsed or visualized. This paper proposes an approach to automatically selecting generation assistance in LLMassisted SysML v2 modeling based on prompt complexity and legacy model hints. For this approach, retrieval augmented generation (RAG) that injects task-specific SysML v2 fragments and domain hints into the prompt, grammar constrained decoding (GCD) based on a lightweight SysML v2 grammar subset, and rule-based post processing (PP) with repair loops are used and compared. The result is a robust approach to engineering assistance that leverages the LLM's existing knowledge, aiming to generate reasonable, syntactically valid SysML v2 Notation. Results indicate that RAG is best used with extensive prompts and detailed product knowledge. At the same time, PP improves syntactic correctness for output close to correct and GCD for short prompts and limited knowledge.
Stats
  • The PDF-Document has been downloaded 25 times.
License/Rightsstatement