Q&A on Any Spreadsheet Requires Interpreting Its Grid Structure
A cell-role annotation framework improves spreadsheet chunking for LLM-driven RAG, but the authors argue discrete classification cannot fully capture 2D grid structure.
The paper proposes splitting spreadsheets into interpretable chunks using semantic cell role annotation, improving chunking interpretability for LLM-driven RAG systems. It beats the state of the art on answer generation through enriched context but hits a hard ceiling, since finite predefined classes cannot capture spreadsheets' continuous relationships and infinite potential cell roles. The authors argue the field must develop dimensionality-reduction techniques that directly flatten 2D unstructured spreadsheets into 1D text for downstream RAG.
- Semantic cell annotation enriches answer generation context, not retrieval accuracy
- Classification models are limited to finite, predefined cell role classes
- Authors propose flattening 2D spreadsheets into 1D text as the path forward
Full article130 words · extracted from arxiv.org · click to collapse
Semantic cell annotation improves chunking interpretability for spreadsheets in LLM-driven RAG systems, aiding answer generation through enriched context rather than improved retrieval accuracy. We propose a novel framework of splitting any spreadsheet into interpretable chunks using cell role annotation. Our framework beats the state of the art, yet it faces a hard ceiling. Spreadsheets are fundamentally two-dimensional unstructured data with continuous relationships and infinite potential cell roles. Because classification models are restricted to finite, pre-defined classes, they cannot perfectly capture this structural nuance, even with human-level annotation. We show that addressing the spreadsheet-to-LLM bottleneck requires moving beyond discrete cell classification. Instead, the field must develop dimensionality-reduction techniques to directly flatten 2D unstructured spreadsheets into 1D unstructured text. Text chunks would be easier for downstream RAG to interpret and generate from.
Text extracted automatically; images, tables and formatting may be missing. Original: https://arxiv.org/abs/2609.20732