From Chatbot to Validator: A Dual-Agent Strategy towards Trustworthy On-Premise Conversational LLMs

Michal Podpora, Marek Baranowski, Aleksandra Kawala-Sterniuk, Mariusz Pelc, Piotr Rogala, Piotr Kawa, Pawel Piróg, Anna Romaniewska, Wojciech Rogala

Abstract: This paper aims to inspire and guide industry leaders and decision-makers in successfully implementing LLM-based agents utilizing additional case-relevant knowledge, beyond LLMs’ original training. The technique presented in this paper, Retrieval-Augmented Generation (RAG), has already proven its practical value in real-world applications. As the RAG offers the injection of use-case-specific knowledge into LLM-based agents, making it possible to design customized expert systems with non-public knowledge. By customizing the expertise area, RAG becomes a powerful tool for businesses to pursue new objectives, and provide new services to their clients. This study presents a structured approach for the Polish Bielik 2.3 model deployment, integrating a Dual-Agent mechanism where one agent manages user interaction and the other ensures response validity and API compliance. The use of embeddings within the retrieval pipeline improves the computational efficiency and response time. The main r

External IDs:dblp:conf/icaart/PodporaBKPRKPRR26