Evaluating Chunking Strategies For Retrieval-Augmented Generation in Oil and Gas Enterprise Documents

Researchers Samuel Taiwo and Mohd Amaluddin Yusoff conducted an empirical study comparing four different chunking strategies for Retrieval-Augmented Generation (RAG) systems when applied to complex oil and gas enterprise documents. Their paper, presented at CCSEIT 2026, tested fixed-size sliding window, recursive, breakpoint-based semantic, and structure-aware methods against a proprietary corpus that included text-heavy manuals, table-heavy specifications, and piping and instrumentation diagrams (P&IDs). The findings demonstrate that structure-aware chunking—which preserves document organization and logical sections—consistently outperformed other approaches in retrieval effectiveness metrics, particularly in top-K accuracy, while also reducing computational costs by approximately 30% compared to semantic chunking methods.

Crucially, the study revealed significant limitations in all text-based RAG approaches when dealing with visually encoded documents like P&IDs, where spatial relationships and graphical elements contain essential information that text extraction misses entirely. This highlights a fundamental gap in current RAG implementations for technical domains where multimodal data is standard. The researchers conclude that while structure preservation is essential for specialized enterprise applications, future RAG systems must integrate computer vision and multimodal language models to properly handle diagrams, schematics, and other non-textual formats common in industrial documentation.