Structured review on rag- and multi-agent frameworks: application-based assessment

Authors: Md Monsur Ali, Abdullah Al Foysal, Siddarth Venkateswaran, Ronald Böck

Abstract:

Multi-Agent Frameworks, Agentic Systems, and related approaches are currently emerging, also in terms of applications. They are mainly based on Large Language Models, especially in terms of conversational agents. In Part I of our contribution, we considered Multi-Agent Frameworks in a theoretical way and provided an overview of state-of-the-art methods. In addition, the current paper presents an application-based assessment of multiple frameworks, covering a wide range of characteristics. As we were interested in the capability, scalability, and complexity of such frameworks in real-world settings, we applied the GenieReader, a tool being designed to address management and extraction of information from complex documents, as application. The frameworks were integrated into the current workflow and tested on data taken from the insurance domain. As a baseline, we used the GenieReader with a Vanilla RAG setting. Finally, we derived multiple hands-on take home messages for the use of Multi-Agent Frameworks.