Description
The introduction of advanced language models like ChatGPT has transformed how we interact with machines and how we interact with artificial intelligence. However, the reliability and accuracy of general knowledge of these models are not without errors, especially when it comes to domain-specific information.
This paper proposes a solution to this challenge by integrating retrieval augmented generation (RAG) into a multi-agent system of diverse large language models (LLMs), namely open source.
This approach allows the creation of knowledge trees, where each agent has specialized knowledge in a specific domain.
The study tested this solution in a realistic case, made available by Demoscore, distributing knowledge among six agents.
The results are promising, demonstrating the ability of LLMs to evaluate and process information that was not used in their training process, and the ability of agents to delegate questions according to their topics and answer questions correctly.
This will make it possible to support collaborators to consult multiple information and documents in a few seconds, contributing to the maintenance support tools of WP13 of the PRODUTECH R3 project.
This publication by the ISEP team, Francisco Oliveira, Luis Gomes, and Zita Vale, will be available in PAAMS 24 proceedings.
This work has been supported by the European Union under the Next Generation EU, through a grant of the Portuguese Republic's Recovery and Resilience Plan (PRR) Partnership Agreement, within the scope of the project PRODUTECH R3 – "Agenda Mobilizadora da Fileira das Tecnologias de Produção para a Reindustrialização".