AI-Driven User Support: A RAG Approach

Information Technology (IT) user support often relies on knowledge scattered across documents inside companies. Finding the right, current answer quickly is hard, and closed-book language models can sound confident while being wrong. This thesis addresses that gap by combining retrieval-augmented generation (RAG) with a light agent layer to produce grounded, auditable assistance. RAG retrieves relevant passages at question time and conditions the response on that evidence, reducing outdated or unsupported claims. On top of this, a simple agent turns the retrieved context into practical outputs such as a clear explanation, a concise ticket, or step-by-step guidance while keeping citations and brief self-checks to maintain faithfulness. The work focuses on what matters most in practice: how documents are chunked, which embedding models are used, how retrieval is tuned, how vector search is configured. Evaluation combines standard retrieval metrics with judgment-based checks of faithfulness and usefulness, alongside operational signals like accuracy. Overall, the thesis offers a straightforward recipe for turning a general Large Language Model (LLM) into a grounded support assistant: retrieve first, write with evidence, and keep outputs simple and ready to act on