Knowledge-Augmented Conversational Agents for Dependency-Aware SysML v2 Model Generation and Refinement

Model-Based Systems Engineering (MBSE) is pivotal for managing the complexity of modern systems. However, applying MBSE to complex products, especially in automatically generating accurate and consistent models in languages such as the Systems Modeling Language version 2 (SysML v2) from requirements, remains a significant challenge, primarily because of the novelty of SysML v2 itself, which is not well-represented in the training corpora of Large Language Models (LLMs), leading to inconsistent or incomplete designs. This paper introduces SysForge, a novel framework that addresses these challenges by transforming the LLM from a simple code generator into a collaborative design partner. Our framework introduces three primary contributions: a structured, comment-enriched sysml knowledge base enabling dependency-aware retrieval-augmented generation; an intermediate structured context synthesis step where an LLM agent refines retrieved information into a coherent summary; and an interactive Designer-Validator conversational loop that ensures correctness and integrity. We present a experiment to evaluate our framework against established baselines, demonstrating its potential to significantly improve the accuracy, completeness, and coherence of automatically generated SysML v2 models.