Enhancing Enterprise RAG Systems through Multi-Agent Architectures: A Case Study in the Insurance Domain

The increasing adoption of Retrieval-Augmented Generation (RAG) architectures by enterprises reflects the growing need to integrate large language models into corporate environments while preserving data confidentiality. RAG enables organizations to harness the advanced capabilities of Natural Language Processing systems without exposing sensitive information to external entities, thus ensuring compliance with privacy and security requirements. This thesis provides a comprehensive analysis of the structure and deployment of a Retrieval-Augmented Generation (RAG) system within an enterprise context. The work begins by detailing the ingestion process of corporate documents, followed by the inference pipeline, and concludes with a systematic approach to automating the entire workflow to achieve a production-ready RAG solution. Building on this foundation, the thesis introduces the design and implementation of an advanced enterprise chatbot based on a RAG architecture, further extended through a multi-agent framework. The proposed approach leverages recent advancements in agent orchestration and graph-based workflows, particularly through the use of libraries such as LangGraph. This integration introduces several innovative functionalities: •??Dynamic user profiling: the system can query internal user databases to deliver personalized responses and recommendations based on individual profiles. •??Customizable guardrails: enabling fine-grained control over the chatbot’s behaviour and compliance constraints within the RAG pipeline. •??Automated topic filtering: improving retrieval efficiency by categorizing and filtering documents according to thematic relevance. •??Cross-topic query handling: ensuring coherent and context-aware responses to complex, multi-domain questions. The proposed solution significantly extends the operational scope of traditional RAG architectures, transforming them from static retrieval systems into adaptive, multi-agent ecosystems capable of performing a broader range of tasks. Experimental evaluations demonstrate improvements in response accuracy, contextual relevance, and system robustness, highlighting the potential of multi-agent RAG architectures as a foundation for next-generation enterprise conversational AI.