Extractive text representations are verbatim segments and sliding windows of text selected directly from a source page by search engines to serve as grounding context for an LLM rather than reading the entire document.