Karpathy shares 'LLM Knowledge Base' architecture that bypasses RAG with an evolving markdown library maintained by AI

TL;DR AI
2 min readKey summary
Andrej Karpathy posted an LLM Knowledge Base design that replaces RAG with a Markdown-based wiki.
The system ingests research materials into a raw/ directory and converts web content to .md using Obsidian Web Clipper.
An LLM compiles the raw files into a structured, linked wiki and runs linting passes to fix inconsistencies.
Community members proposed scaling the approach with multi-agent orchestration and validation models for enterprise use.
