Agentni sistem RAG za podporo zaposlenim v večjih podjetjih

We present the design, implementation, and evaluation of the RAG agent system, which enables employees of a large telecommunications company to conversationally search for information in a knowledge base built from the content of a public website. The system consists of a data capture pipeline that converts web content into vector entries in the Azure AI Search database, and an agent application built with the LlamaIndex framework that, following the ReAct model, decides on a hybrid search itself and supports each answer with a reference to the source. In addition to the cloud language model, the system uses locally hosted models via the Bifrost AI gateway, in which the data does not leave the company's infrastructure. In a set of 105 real questions from test users, the share of successful answers according to a broader definition, which also includes partially correct and justified rejected answers, is 94.3%, and the share of fully correct answers is 67.6%; the user usually receives the first part of the answer after about five seconds. A survey among test users confirms the time savings, but highlights the coverage of the knowledge base and confidence in the answers as the main challenges.