I have successfully rendered my entire production codebase as a knowledge graph.
The technical details include:
- 293 source files
- 5,638 nodes (functions, classes, routes)
- 7,495 edges (calls, imports, inheritance)
- 400 communities clustered by Leiden, with EXTRACTED / INFERRED / AMBIGUOUS provenance labels on every edge, ensuring transparency.
- Built from tree-sitter AST with zero LLM cost and rebuilt on every push.
In plain English, every piece of the system has a name, an address, and a precise relationship to every other piece. Clicking any node reveals what it interacts with, and searching any term zooms the graph to that specific point.
What this unlocks:
Engineers can query the graph instead of relying on grepping. For example, "What calls fetch_ai_visibility?" now returns 5 nodes instead of 5,000 lines of output.
AI agents utilize it as bounded context, resulting in the same Claude session producing 50 times fewer tokens and much sharper answers.
Non-technical teammates and operators can finally visualize the actual structure of what they're working on, rather than having to take it on faith.
Credits: https://lnkd.in/e-n3ZM2q
