Since this transition presents significant challenges, consulting firms are playing a crucial role in supporting businesses navigate this shift, using AI-driven solutions to enable informed decisions. However, the standard use of large language models (LLMs) in chatbot systems often leads to problems such as hallucinations and inaccurate responses.

To address this, retrieval-augmented generation (RAG) approaches improve response accuracy by incorporating a vector store that provides relevant context to the language model. As a result of this new technology, this master’s thesis investigates the optimal design for a retrieval-augmented chatbot architecture when answering Industry 4.0 action recommendation questions by assessing the performance of various LLMs in conjunction with topK and temperature parameters.

The analysis shows that OpenAI's GPT-4o model, when used with a low temperature setting and a high topK value within a Sequential Agentic Pattern architecture, provides the best results when answering Industry 4.0 action recommendation questions. This configuration demonstrates superior performance in terms of response completeness and clarity compared to other LLM model (Gemini 1.5-Flash, Claude-3.5-Sonnet, Mistral-Large 2402) and parameter combinations. Furthermore, this thesis provides a comprehensive framework for architecture design, document upload and chatbot evaluation, making the results publicly available for further research and practical applications.