Entwicklung von KI Assistenten basierend auf RAG

Large Language Models (LLMs) have demonstrated strong capabilities in natural language understanding and generation, yet their effective integration into enterprise workflows remains challenging. In practice, organizations face difficulties related to accessing company-specific knowledge, enforcing task boundaries, and aligning LLM-based systems with existing operational processes. In particular, unstructured conversational assistants are insufficient for knowledge-intensive domains such as IT support, where reliability, traceability, and process integration are essential.

This Studienarbeit investigates the design and implementation of a configurable platform for retrieval-augmented, workflow-driven AI assistants in enterprise environments. Building upon an existing secure and scalable RAG-as-a-Service Infrastructure. The system is extended with a modular tool integration mechanism, a multi-stage workflow engine, and a flexible data ingestion pipeline. These components enable the controlled execution of structured tasks, dynamic access to external systems, and continuous synchronization of organizational knowledge sources.

The proposed architecture is evaluated through a proof-of-concept implementation, focusing on a Level~1 IT support scenario involving guided information gathering and automated ticket creation. The results indicate that constraining LLM behavior through explicit workflows and retrieval scopes improves task consistency and operational applicability compared to generic conversational approaches. While the implementation represents a proof of concept rather than a production-ready system, this work contributes architectural patterns and design principles for integrating LLMs into enterprise workflows in a controlled, extensible, and configurable manner.