Karpathy Shares LLM Knowledge Base Architecture: RAG-Free Personal Wikis at 400K Words via Obsidian + Claude

VentureBeat·high signal

Andrej Karpathy posted a detailed workflow for using LLMs to build personal knowledge bases — dumping source documents (papers, repos, articles) into a raw/ directory, then having an LLM 'compile' structured markdown wikis with backlinks, summaries, and categorization. Obsidian serves as the read layer; the LLM writes and maintains all content. At scale (~100 articles, ~400K words), Karpathy runs 'health checks' where the LLM finds inconsistencies, imputes missing data via web search, and surfaces candidates for new articles. The approach bypasses RAG entirely with an evolving, LLM-maintained markdown library.