This article examines how semantic cell annotation improves chunk interpretability in large‑language‑model‑driven retrieval‑augmented generation (RAG) systems for spreadsheets. By assigning each cell a role—such as header, value, or formula—we can split any two‑dimensional spreadsheet into understandable chunks. Experiments show that this approach enriches the context for answer generation, offering better generation quality than prior methods, even though retrieval accuracy sees little change.
Spreadsheets are fundamentally 2D unstructured data with continuous relationships and potentially infinite cell roles. Classification models are limited to a finite set of predefined classes, so even with human‑level annotation they cannot fully capture these structural nuances, creating a hard performance ceiling.
To overcome the spreadsheet‑to‑LLM bottleneck, we advocate moving beyond discrete cell classification toward dimensionality‑reduction techniques. The idea is to flatten the 2D spreadsheet directly into a 1D text sequence, making it easier for downstream RAG components to interpret and generate. Compared with role‑based chunking, pure text chunks may provide better contextual coherence and generation efficiency.
Review