Lifecycle-aware knowledge graphs with human-in-the-loop agentic retrieval for automotive manufacturing

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Modern automotive manufacturing demands agile coordination; however, operational decision-making remains severely hindered by data fragmentation across siloed enterprise systems and the lack of intuitive information retrieval mechanisms. While graph-based models offer a promising foundation for structural data integration, querying them requires formal languages (e.g., Cypher) that remain largely inaccessible to domain experts. To address these challenges, this paper proposes a lifecycle-aware knowledge graph schema and an agentic retrieval framework for cross-domain integration and retrieval in automotive manufacturing. First, we introduce a multi-dimensional graph schema that unifies automotive manufacturing knowledge across product design, process planning, production organization, and shop-floor execution. To faithfully capture the complexity of such heterogeneous data, the schema explicitly separates manufacturing entities into four complementary dimensions: persistent identity, engineering specification, released version, and runtime state. Second, we develop a schema-grounded, complexity-adaptive agentic retrieval framework designed for condition-based structural retrieval to enable reliable and flexible access to domain knowledge. By integrating Large Language Models (LLMs) reasoning with targeted human-in-the-loop confirmation gates, the framework translates natural language queries into verifiable and executable subgraph matching and path traversals. The proposed approach is validated using real-world data sources. Structural evaluations demonstrate its capability to support cross-domain integration and multi-hop query traversal, while quantitative experiments show that the agentic retrieval pipeline significantly outperforms conventional generative baselines, achieving an execution accuracy of 95.0% on a rigorous industrial benchmark. Overall, this work provides a structured foundation for lifecycle-aware knowledge representation and reliable graph-based retrieval in automotive manufacturing.