Chatbot technické podpory

Abstract

Large language models bring significant potential to the field of technical support, yet they suffer from fundamental limitations -- a fixed knowledge cutoff and a tendency to generate convincingly phrased but factually incorrect responses. This bachelor's thesis presents the design and implementation of a technical support chatbot that overcomes these limitations using the Retrieval-Augmented Generation (RAG) architecture.

The proposed system is implemented as a .NET 10 Web API and uses Azure Cosmos DB as a vector database for storing documentation embeddings. The query processing pipeline includes query rewriting using an LLM, semantic search, and result reranking using the Jina Reranker v3 cross-encoder model. Orchestration of language model calls is handled through the Microsoft Semantic Kernel library.

The system was evaluated on a set of 21 queries divided into five categories covering factual, procedural, dependency-based, and conversational queries. The system achieved an overall success rate of 76\% and none of the responses were evaluated as factually incorrect -- the system never hallucinated. The results confirm that the RAG architecture combined with a well-designed system prompt effectively eliminates hallucinations and enables the chatbot to respond accurately based on up-to-date documentation.