The limited availability and flexibility of non-cloud solutions manifests a gap in the market. Companies working with sensitive data require an offline solution for a precise and adaptable product. This paper establishes a concept that combines state-of-the-art technologies with a configurable core.

First, foundational knowledge about natural language processing had to be gathered. Three main sources were studied to establish a strong foundation on the topic. Christopher D. Manning from Standford University has shared all lectures from the CS224N class on YouTube. Similarly, Ryan Cotterell from ETH Zürich has published his lectures on NLP on his personal website. Finally, the book "Introduction to Natural Language Processing" by Jacob Eisenstein served as a tool for guidance. Together with the supervising lecturer and the project team, a possible system architecture was designed.

A prototype was developed that abstracts the functionalities of a system and makes them available through a chatbot interface. Whereas Large Language Models seem incredibly smart, the result is statistical nature, and the technology does not grasp a real understanding. It can be assumed that machine learning from text only is too limited and cannot be trusted with critical tasks yet. Creating a defined space for the chatbot does not require a full understanding of the situation but only how to sort the incoming requests into the defined buckets. The limitation of space benefits the chatbot enormously, and it attains a better focus on a specific task. Furthermore, the configuration of the buckets can easily be changed and used for different systems. The prototype introduces a certainty threshold to further optimize the system with more samples that have not been recognized the first time. While this first version confirms a solid system architecture, more data should be gathered or created for a more precise recognition. The system can be introduced as an alpha version to collect data for future research.