Research on Retrieval-Augmented Generation Methods Based on Agent Workflows and Applications

Retrieval-Augmented Generation(RAG) is a technique that enables large language models to leverage external knowledge sources for generation. However, current RAG systems are either text-only, which prevents the utilization of visual information such as layout and images, or employ fixed chunk sizes, adversely affecting retrieval performance. In this study, we propose an AgentWorkflow-based RAG chunking method that utilizes the reasoning capabilities of LLMs to dynamically adjust chunks based on the required external data types and task priorities. Experiments demonstrate that AgentWorkflow outperforms traditional text-based RAG in both retrieval and generation phases, achieving over a 3% end-to-end performance improvement compared to baseline RAG solutions.