All videos
0:00 / 0:00
research

Inside DeepWiki: How Cognition Builds Wikis for Devin at Scale

LangChain21 August 2026Watch on YouTube

Description

Jacob Teo from Cognition's Research and Engineering team breaks down DeepWiki, the auto-generated codebase documentation product behind Devin, at a LangChain SF meetup. He covers how DeepWiki's wiki algorithm evolved from a heavily orchestrated v1 to a more agentic v2, the quantitative signals Cognition uses to judge wiki quality, and four context engineering principles for deciding what belongs in a wiki versus what an agent should find on its own. Chapters: 0:00 What is DeepWiki? 0:26 The origin story: DeepWiki started as context for Devin 1:05 From internal tool to 1.4 million indexed repos 2:27 The DeepWiki algorithm and why some orchestration still matters 6:07 How DeepWiki actually builds a wiki, step by step 7:17 What a codebase graph looks like at scale 7:48 Upgrading the wiki algorithm from v1 to v2 9:30 How Cognition measures whether a wiki is good enough 11:33 Shifting to wikis for agents and codebase intelligence 11:50 RAG versus agentic search: why it isn't a dichotomy 12:40 Four context engineering principles 14:04 Path compression versus unknown unknowns 15:36 What's still unsolved in codebase intelligence Resources: → DeepWiki: https://deepwiki.com → DeepWiki on GitHub: https://github.com/CognitionAI/deepwiki → LangGraph: https://www.langchain.com/langgraph → LangSmith: https://www.langchain.com/langsmith

What you'll learn

  • DeepWiki is an AI system that automatically generates codebase documentation, originally developed as a context source for Devin.
  • DeepWiki's algorithm evolved from a heavily orchestrated version 1 to a more agentic version 2 with greater autonomy.
  • Cognition uses quantitative signals to measure wiki quality, correlating with how well agents perform their tasks.
  • Context engineering follows four principles for deciding what belongs in a wiki versus what agents should discover themselves.
  • RAG and agentic search are not mutually exclusive, but can both be used within a context engineering strategy.

Frequently asked questions

What is DeepWiki and what was it originally built for?
DeepWiki is a system that automatically generates codebase documentation. It was originally developed as an internal context source for Devin, Cognition's AI assistant, and has grown into a product that indexes 1.4 million repositories.
How does the DeepWiki v2 algorithm differ from version 1?
Version 1 was heavily orchestrated with significant manual control, while version 2 is more agentic and gives agents greater autonomy. Despite this shift, some orchestration still proves valuable.
What are the four context engineering principles for wikis?
The video presents four context engineering principles for deciding what belongs in a wiki versus what agents should discover themselves. These principles help optimize what is pre-documented versus what agents should find dynamically.
How does Cognition measure if a wiki is good enough?
Cognition uses quantitative signals and metrics to assess wiki quality. These signals directly correlate with how well agents perform their tasks.

Topics

ai agentscodebase intelligenceragcontext engineeringdevin