Quick Facts
- Andrej Karpathy developed a three-stage LLM knowledge base system that processes raw data into self-maintained markdown wikis
- The approach handles 100 articles and 400,000 words without traditional RAG infrastructure or vector databases
- System enables full traceability where every AI claim links back to human-readable markdown files
Andrej Karpathy, former Director of AI at Tesla and OpenAI co-founder, revealed a new architecture for LLM knowledge management that bypasses traditional RAG systems entirely. The approach uses AI to compile raw data into self-maintaining markdown libraries.
The three-stage system starts with data ingestion, where research papers, GitHub repositories, datasets, and web articles dump into a raw directory. Karpathy uses Obsidian Web Clipper to convert web content into markdown files, storing images locally for LLM vision access.
The core innovation happens in the compilation stage. Instead of indexing files, the LLM reads raw data and writes structured wikis with summaries, key concepts, encyclopedia-style articles, and backlinks between related ideas. The system then runs health checks where the LLM scans for inconsistencies and missing connections.
Karpathy’s example handles roughly 100 articles and 400,000 words on research topics. He expected to need complex RAG pipelines but found the LLM maintains index files and document summaries effectively at this scale.
The architecture solves the “black box” problem of vector embeddings by making markdown files the source of truth. Every AI claim traces back to specific files humans can read, edit, or delete. This eliminates retrieval noise common in traditional RAG systems.
Tech podcaster Lex Fridman confirmed using a similar setup, adding dynamic HTML generation with JavaScript for interactive data sorting and visualization. “I have the system generate a temporary focused mini-knowledge-base that I then load into an LLM for voice-mode interaction on long runs,” Fridman wrote.
The business implications attracted immediate attention. Entrepreneur Vamshi Reddy noted: “Every business has a raw directory. Nobody’s ever compiled it. That’s the product.” Karpathy agreed, calling it an “incredible new product” category.
Most companies struggle with unstructured data across Slack logs, internal wikis, and PDF reports. A Karpathy-style enterprise system wouldn’t just search documents but would author a company knowledge base that updates in real-time.
McKinsey research shows employees spend 1.8 hours daily searching for information, totaling 9.3 hours weekly. This represents one-fifth of working hours devoted to resource hunting rather than productive tasks.
The system addresses the core frustration of stateless AI development where context resets force developers to reconstruct project understanding. By maintaining persistent, LLM-maintained records, the architecture preserves institutional knowledge across sessions.
Karpathy currently runs the system as “a hacky collection of scripts” but sees potential for enterprise products. The approach represents a shift from LLMs as answer machines to knowledge infrastructure that actively maintains organizational memory.