Inquire, Extract, Retrieve and Re-rank: A Multi-Agent Approach for Conversational Recommender Systems

Abstract

The overwhelming abundance of options on digital platforms presents significant challenges for users seeking personalized recommendations. Traditional recommendation systems, often reliant on historical user data and collaborative filtering, may not accurately capture current user preferences and can lack consistency and reliability, especially in cold-start scenarios. With the advent of Large Language Models (LLMs) capable of engaging in dynamic conversations, Conversational Recommender Systems (CRS) have emerged as a promising solution to enhance personalization through natural language interactions. We introduce a multi-agent CRS framework that leverages LLMs to provide personalized game recommendations without relying on historical data. Our system comprises five specialized agents: an Inquiry Agent to elicit user preferences, an Extractor Agent to structure these preferences into a user profile, a Semantic Retriever to semantically match preferences with game attributes, and two Re-rankers to refine the recommendations based on relevance and quality. We apply our framework to a comprehensive dataset of over 97,000 video games from Steam and employ synthetic personas generated by LLMs to simulate user interactions. The experimental results demonstrate the system's potential in effectively capturing user preferences and delivering tailored recommendations despite challenges associated with evaluating performance using synthetic data. By advancing these AI-driven approaches, this work lays a foundation for more intuitive and responsive recommendation platforms and search engines. It transforms how users interact with digital environments by addressing the cold-start problem in recommendation systems, paving the way for enhanced user experiences through personalized, dialogue-based interactions.