Karpathy's LLM Wiki Pattern

SmartDocs is built on a key insight from Andrej Karpathy's LLM Wiki pattern (April 2026, 5,000+ stars).

The Core Idea

Most AI + documents experiences work like this:

Upload files → ask question → LLM retrieves chunks → generates answer

The problem: the LLM re-derives understanding from scratch on every query. There's no accumulation. Ask a subtle question requiring synthesis across five documents, and the LLM finds and pieces together fragments every single time.

Karpathy's alternative: build a persistent, interlinked wiki that compounds over time.

"The knowledge is compiled once and then kept current, not re-derived on every query. This is the key difference: the wiki is a persistent, compounding artifact."

How SmartDocs Applies This

Your documentation is already a compiled artifact — you wrote it. SmartDocs reads the connections between your pages and uses them to provide better AI answers.

The Three Layers

Karpathy's architecture defines three layers:

Layer	In Karpathy's Wiki	In SmartDocs
Raw sources	Articles, papers, PDFs — immutable	Your `content/*.md` files — immutable
The wiki	LLM-generated markdown pages, interlinked	Already written by you. SmartDocs reads the links.
The schema	CLAUDE.md telling the LLM how to structure	`build-graph.js` + `page-index.json`

The Key Difference

Karpathy's pattern uses an LLM agent to write the wiki. SmartDocs doesn't need to — you already wrote it. SmartDocs does what the LLM agent would do: read your cross-references and build the graph.

Operations: Ingest, Query, Lint

Karpathy describes three operations for maintaining a wiki:

Ingest — Add a new source. The LLM reads it, extracts key info, updates entity pages, flags contradictions, updates the index. One source can touch 10-15 wiki pages.
Query — Ask a question. The LLM finds relevant pages, reads them, synthesizes with citations. Good answers get filed back into the wiki.
Lint — Health check. Look for contradictions, stale claims, orphan pages, missing cross-references.

SmartDocs automates the Query operation. When a reader asks a question, SmartDocs finds the right pages, follows their links, and feeds full context to the LLM. The answer comes back grounded in your documentation.

The Lint operation is something you can do manually — check that your pages link to each other where relevant. Every link becomes a graph edge that improves retrieval.

Comparison with Traditional RAG

	Traditional RAG	Karpathy LLM Wiki	SmartDocs
Knowledge persists?	No — re-derived each query	Yes — compiled once	Yes — your docs
Cross-references?	No — disconnected chunks	Yes — interlinked pages	Yes — built from your links
Needs embedding API?	Yes	No (index file)	No (index file)
Needs vector DB?	Yes	No (markdown files)	No (JSON graph)
Hallucination risk	High (fragments lack context)	Low (full pages)	Low (full pages)