Recommender systems can be found in every app, website and service today, as they are essential for guiding users through the massive amount of information and data they’re exposed to. Recently, large language models (LLMs) have emerged and experienced widespread adoption. This has led to a shift towards LLM-based recommender systems, as LLMs can provide a more human-like recommendation experience. Therefore, we propose RAG-Rec, a tool that aims to support an LLM or generate recommendations independently. RAG-Rec is a retrieval-augmented generation system in which the retriever searches a knowledge base for users similar to the target user. The retriever then returns a set of items liked by similar users, which can be used for recommendations or as input to an LLM. The LLM will then use its reasoning abilities to generate a more refined recommendation list. We tested our model on the MovieLens 1M dataset, and included metrics for recall, precision, F-score, and hit rate. We combine our retriever with GPT4o and GPT4o mini, and we also provide metrics for both LLMs without the retriever as a baseline. We found that our proposed retriever outperformed the Retriever + LLM and LLM-only setups, with a hit rate@40 of 0.9942 and a recall of 0.43764.
Aly et al. (Thu,) studied this question.