RAG Strategies for a Hallucination-Free University Admission AI Assistant.

The development of a LLM-based university admission assistant requires the use of up-to-date and factually correct information. The state of the art proprietary and open source models are prone to hallucination (generating false information because of a lack of specific real information in their pretraining phase), which might cause confusion for the prospective students. A set of retrieval augmented generation (RAG) strategies on different models are evaluated, that ensure that the models output up -to-date and factually correct information and are also able to use personalized data to help the prospective students with the application process. We are using a specific case for UNWE, Sofia, which also adds an additional requirement for the support of Bulgarian language in the conversation. [ABSTRACT FROM AUTHOR]