Advanced AI-Driven Solutions for Enhancing Contract Management: Multi-Agent Systems, Bi-Modal Models, and Adaptive Retrieval

Contract management in Revenue Accounting and Reporting (RAR) systems within Enterprise Resource Planning (ERP) presents challenges due to the complexity and volume of documents. This thesis introduces a multi-agent "Agent Graph" interface with drag-and-drop functionality, allowing users to easily create complex, task-specific chatbots by connecting agents and functions, streamlining automation in contract management. Additionally, we evaluate multiple small language models on the HotpotQA dataset within a specialized multi-agent setup, achieving up to 75% match accuracy. Following this evaluation, the models are applied to contract management tasks, such as metadata tagging and anonymization. A fast bi-modal system further reduces processing times by up to 88% for intensive document grounding, enabling efficient handling of both textual and visual contract data without relying on resource-heavy models like GPT-4o. Finally, an Adaptive User-Driven Retrieval-Augmented Generation (AUD-RAG) system is implemented, achieving 89.97% NDCG and 88.15% accuracy with GPT-4o through human feedback-driven weight adjustment. These methodologies offer scalable, AI-driven solutions that significantly improve the efficiency and reliability of contract management in ERP systems.