卡帕西力推的 LLM Wiki，会淘汰传统 RAG 吗？

TL;DR - LLM Wiki precompiles source documents into agent-oriented, structured Markdown knowledge bases, reducing repeated retrieval and reasoning costs. It complements rather than replaces RAG, especially at larger scales or when source-level detail matters.

• Unlike query-time RAG, LLM Wiki performs summarization, classification, and semantic linking during document ingestion.
• Cognition, Factory, LangChain, and GBrain have implemented variants for code documentation, personal knowledge, and agent retrieval.
• Pure wiki navigation suits roughly 100 sources; larger collections require hybrid keyword, vector, and LLM-ranked retrieval.
• Wikis risk lossy summaries, stale content, and wasted ingestion costs, and they do not replace user-specific interaction memory.