Why LLM Wiki

The problem

You have code, docs, papers, screenshots. You want to ask questions about them.

Approach	Issue
Paste into chat	Hits context limits, no persistence
RAG pipeline	Complex setup, vector DB, chunking, embedding model
Full re-read	Expensive, slow, no cross-document connections

None compile knowledge once and let you query forever.

Karpathy’s three layers

“A large fraction of my recent token throughput has gone not into manipulating code, but into manipulating knowledge.”

Raw — your files, never modified
Compile — structured knowledge with cross-references
Query — ask questions without re-reading source files

Compilation is separate from querying. Build the graph once, query for weeks.

Graphs vs. embeddings

	Vector DB (RAG)	Knowledge Graph
Setup	Embedding model + chunking + DB	`pip install my-llm-wiki`
Cross-references	Chunks are isolated	Edges connect everything
Structure	Flat similarity	Communities, paths, hubs
Explainability	“These chunks are similar”	“A calls B which imports C”
Persistence	Running DB	JSON file on disk
Cost	Embedding API calls	Free (AST) or one-time (agent)

At personal scale (10-1000 files), you don’t need approximate nearest neighbor. You need what connects to what.

The living wiki

Karpathy’s vision goes beyond “compile once” — the wiki is a persistent, compounding artifact. Every session adds knowledge. Insights get filed back. The graph grows over time.

Monitor → Rebuild → Lint → Write-back → Report
   ↑                                       │
   └───────────────────────────────────────┘

Monitor — detect file changes, trigger rebuild
Rebuild — SHA256 cache skips unchanged files
Lint — find orphans, tiny communities, high ambiguity
Write-back — file insights as markdown, merge into graph
Report — track growth: nodes, edges, communities over time

This cycle is implemented in the Claude Code skill. The agent follows it automatically when you run /wiki ..

Three confidence levels

Every edge is tagged:

EXTRACTED — found directly in source (import, link, citation)
INFERRED — reasoned by extractor (shared concept, implied dependency)
AMBIGUOUS — uncertain, flagged for review

You always know what was found vs. guessed.