Dynamic Multilingual Retrieval-Augmented Generation (RAG) System for Table-Heavy PDFs

Abstract

Baseline Retrieval-Augmented Generation (RAG) systems show great promise in helping AI generate relevant and grounded responses while reducing hallucinations and lowering computational costs [1][2]. By merging information retrieval with large language models (LLMs), RAG frameworks let models access external knowledge sources, which improves factual accuracy and response reliability. However, baseline RAG systems mainly work with unstructured sequential text and struggle with structured data formats, especially tables [3][4]. Tabular data often has complex relationships between rows, columns, headers, and numerical values. Conventional text-based retrieval methods do not capture these relationships well.

In real-world situations, many documents, like research papers, financial reports, and enterprise PDFs, mix unstructured text with structured elements such as tables, charts, and multi-level headers. Traditional RAG pipelines often fail to keep the two-dimensional structure of tables during preprocessing. This leads to fragmented or misleading retrieval results. Such structural issues can harm downstream generation, causing incomplete or inaccurate responses, especially when numerical or table-related information is involved.

To tackle these challenges, this study investigates a better way to retrieve complex multi-format data in a format friendly to LLMs [5][6]. The suggested method emphasizes structure-aware preprocessing and semantic chunking strategies that maintain contextual and structural relationships in both text and tables. By creating semantically meaningful chunks with controlled overlap, the approach ensures that essential contextual information stays intact during retrieval.

Additionally, the system supports table-heavy documents written in multiple Latin-based languages by using multilingual embedding models. This allows for both same-language and cross-language retrieval, enabling users to query documents in different languages without losing semantic accuracy. Overall, the proposed method improves the performance of RAG systems for complex, real-world document collections by enhancing retrieval accuracy, contextual grounding, and usability across various document formats.