This paper proposes a Retrieval-Augmented Generation (RAG)-based question-answering (QA) system to address the challenges of information access due to the vast volume and unstructured nature of internal corporate documents. Using crash test documents from the automotive industry as an example, we focus on processing diverse data types, maintaining data confidentiality, and ensuring traceability between generated answers and the original documents. The proposed system consists of a data pipeline that transforms various document types into a structured corpus and QA pairs, an on-premises privacy-preserving architecture, and a lightweight reference matcher that links answers to supporting content. Application to the automotive industry demonstrates improvements in factual accuracy, informativeness, and usability compared to existing systems.
