A modular tool-driven RAG framework for engineering test documentation

Engineering test procedures encapsulate decades of domain knowledge and experiential decision-making accumulated by human experts. However, reusing this knowledge across projects remains labour-intensive, as engineers must repeatedly interpret historical documents and customer-specific requirements. This study proposes a design-oriented, tool-driven framework for procedural knowledge reuse, integrating retrieval-augmented generation (RAG) with structured parameter handling and template-guided generation. The system adopts a single-agent architecture that coordinates evidence retrieval, constraint alignment, and controlled content generation, enabling the explicit use of specification conditions in procedure drafting. By grounding generation in retrieved evidence, the framework supports consistency, traceability, and alignment with engineering standards while reducing reliance on model-internal knowledge. A case study in transformer testing shows that the proposed approach achieves a condition-detection accuracy of 0.969, a BERTScore F1 of 0.92, and reduces average drafting time from 52 to 15 min. These results suggest its potential applicability to documentation tasks with similar procedural characteristics.