Inside DeepWiki: How Cognition Builds Wikis for Devin at Scale
Cognition's DeepWiki has indexed 1.4M GitHub repos and powers Devin's context layer
“the hardest part of creating a wiki is actually the table of contents”— Andrej Karpathy
Cognition's DeepWiki, originally built as an internal context layer for their Devin coding agent, has scaled to 1.4M indexed GitHub repositories and 20M served queries. The system uses heuristics and codebase graphs—incorporating directory structure, symbol graphs, git history, and runtime data—rather than pure LLM orchestration, to handle enterprise codebases up to 20GB. The core engineering insight is that table-of-contents generation is the highest-leverage step: a wrong ToC corrupts the entire wiki regardless of page-level writing quality.