A RAG-based Large Language Model Framework for Tracing Requirements to Design Information of Automated Driving Systems

The development of Automated Driving Systems (ADS) entails managing hierarchical architectural views in which driving features are systematically specified, implemented, verified and validated with heterogeneous design artifacts. Due to the safety-critical nature of such systems, one of the most important development tasks is to systematically manage the system requirements across multiple abstraction levels, ranging from high-level system features to detailed technical specifications and corresponding verification and validation test cases. Recent advances in Large Language Models (LLM) show limited applicability in such system development context, as models trained on large-scale open-domain corpora often lack the precious, domain-specific reasoning required for requirement analysis beyond general linguistic competence. To alleviate these limitations in requirement analysis of ADS, this paper presents a functional framework that integrates a toolchain combining system models and LLM to trace requirements to design information through Retrieval-Augmented Generation (RAG)–based solutions. Specifically, we propose a knowledge base by modeling the design information of ADS. The knowledge base is constructed by graph-based models, supporting an interpretable process of multi-depth retrieval and reasoning. As a result, the LLM makes precious responses by integrating the input requirements and the retrieved domain knowledge. To evaluate the proposed framework, we create a dataset by synthesizing design information from public resources. In addition, we use a more extensive public dataset to assess the performance of the framework in terms of robustness and adaptiveness. The results indicate the proposed framework shows significant improvement over the baseline methods.