Recomendação conversacional de produtos em comércio eletrônico: uma abordagem baseada em agente LLM e busca vetorial sobre representação tokenizada

Abstract

Traditional search systems in e-commerce rely on structured filters that require users to translate their needs into categorized attributes, distancing themselves from the natural way consumers express their purchasing desires in free language. Large Language Models have the potential to reduce this gap, but their isolated use in recommendation scenarios faces relevant limitations, particularly regarding access to domain-specific and up-to-date information, and the tendency toward hallucination. This work investigates how LLM-based agents, equipped with external tools, can be designed to perform conversational product recommendation from requests expressed in natural language. A chatbot based on an LLM agent was developed, operating on a tokenized representation of products, performing vector similarity search over an index hosted on the Databricks platform, and conducting preference elicitation to handle underspecified requests. The experimental scope was restricted to four categories of consumer electronics, with user queries conducted in Brazilian Portuguese. The experimental evaluation, conducted on 15 representative test cases, indicated, within the evaluated scope, adequate correspondence between the agent’s behavior and the expected outcomes across all considered dimensions, with no occurrence of factually unsupported responses. The over-elicitation phenomenon and the strategy adopted for its mitigation were also documented. The results suggest that the proposed architecture constitutes a promising approach for conversational recommender systems in e-commerce.