文章中的相关论述

"When an agent needs information from your codebase or document library, it typically does the equivalent of a project-wide Ctrl+F. Every matching line is returned — unranked, unfiltered, and unprioritized."

"Irrelevant matches pile into the LLM's context window, forcing the model to read and process thousands of tokens it didn't actually need."

"QMD (Query Markup Documents), built by Shopify founder Tobi Lütke, adds a third stage: LLM re-ranking. After BM25 and vector search each return candidates, a local language model re-reads the top results and reorders them by actual relevance to your query."

"The token costs that come with always-on AI agents aren't inevitable. This guide covered two areas where better tooling can cut the waste: search and memory."

"index1 and QMD solve this on a single machine by combining BM25 keyword scoring with vector search and returning only the most relevant results."