Developing a RAG System for R&D Automation in Reka Rubber’s Manufacturing Supply Chain Process

Abstract

This thesis describes the design, development, and evaluation of a Local Retrieval (RAG) additive manufacturing system for the R&D department of Reka Rubber. The project aimed to automate document retrieval and management while ensuring complete data privacy within Reka Rubber’s internal network. The system, RAG, implemented in Python, brings together LangChain, SentenceTransformers, FAISS, and Ollama on a Debian-based server for secure and offline inference of large language models (LLMs). A user interface built with Streamlit was created to serve R&D engineers, a target user group without a technical background. The prototype was tested using R&D documents, such as product certificates and test compliance reports, and compared with the JollaMind2 business assistant. Quantitative results showed that retrieval accuracy and the number of irrelevant outputs were higher (0.80 vs. 0.78, respectively), and qualitative feedback confirmed less manual verification and greater readiness for compliance. The RAG assistant proved an effective tool for meeting GDPR requirements, improving document workflow compliance, and enabling scalability, though batch-processing capabilities remained limited and hardware constraints persisted. Research has indicated that local AI is still capable of digitally transforming businesses and improving operational stability, making this a strong move towards the adoption of secure AI in industrial environments.