Mark David Longar, Jakob Fir and Bor Pangeršič

Abstract
The rapid growth of video streaming platforms has intensified
the demand for personalized content recommendations. How-
ever, current solutions often rely on historical user data, leading
to challenges like the cold start problem and overlooking users’
immediate preferences. We present a conversational recommen-
dation system that leverages large language models (LLMs) to
generate keyword-based content and query descriptions. By in-
tegrating Retrieval-Augmented Generation (RAG), our system
efficiently retrieves relevant content, independent of prior user in-
teractions, and ensures consistent performance across languages.
Preliminary testing shows our system outperforms the RAG base-
line by up to 24% in less descriptive queries and demonstrates
consistent performance across three languages. While the results
are promising, further evaluation focusing on user interaction
and satisfaction is necessary. Our approach can potentially be
extended to other recommendation systems, offering broader
applicability and enhanced content personalization.