Government product recommendation systems in e-Katalog: leveraging large language models with retrieval-augmented generation

Indonesia's electronic catalog (e-Katalog) is a platform developed to facilitate government procurement, providing an extensive collection of products and services that simplifies purchasing processes, enhances transparency, and reduces costs. However, the exponential growth in the number of listed products has led to challenges such as information overload, complicating the selection process for buyers. E-katalog lacks the ability to effectively deliver personalized, context-aware recommendations that could further support procurement strategies. To address these challenges, this research aims to develop a product recommendation system for e-Katalog using large language models (LLMs) integrated with the Retrieval-Augmented Generation (RAG) framework. The goal is to provide contextual and relevant recommendations, prioritizing MSMEs and local products to support effective procurement decision-making by government institutions. This study compares three embedding models all-MiniLM-L6-v2, all-distilroberta-v1, and all-mpnet-base-v2, based on their effectiveness in product retrieval, with all-mpnet-base-v2 achieving the highest average similarity score of 0.72. This enabled the RAG-based system, powered by GPT-3.5 Turbo, to deliver more targeted and meaningful product suggestions, thereby effectively promoting local products and MSMEs.