Evaluating RAG Accuracy Enhancement via Page Merging for Non-Ideal Documents

It is known that Retrieval-Augmented Generation (RAG) systems employing less capable Large Language Models (LLMs), especially in local environments, have low inference accuracy when dealing with large tables and figures that extend across multiple pages in administrative documents. This issue arises because page breaks disrupt the association between data elements that should ideally be linked on a single page. To tackle this problem, we introduce “PDF Merging”, a method that involves a layout-aware page merging process applied beforehand. Our empirical evaluation was performed on Japanese administrative documents featuring many separate tables. The findings indicate that PDF Merging enhances accuracy, provided the embedding model performs adequately. Notably, we discovered that selecting the Markdown format for table extraction affects the outcome substantially.