김경종
3.8 KiB
type, title, complexity, domain, aliases, created, updated, tags, status, related, sources
| type | title | complexity | domain | aliases | created | updated | tags | status | related | sources | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| concept | LLM Wiki Pattern | intermediate | knowledge-management |
|
2026-04-07 | 2026-04-07 |
|
mature |
|
LLM Wiki Pattern
A pattern for building persistent, compounding knowledge bases using LLMs. Originated by Andrej Karpathy. The key insight: instead of re-deriving knowledge from raw documents on every query (RAG), the LLM incrementally builds and maintains a structured wiki that gets richer with every source added.
The Core Idea
Most AI knowledge tools work like RAG: index raw documents, retrieve chunks at query time, generate an answer. Nothing accumulates. Ask a question that needs five documents and the LLM reassembles fragments every time.
The wiki pattern is different. When a new source arrives, the LLM reads it, extracts what matters, and integrates it into the wiki: updating entity pages, noting contradictions, strengthening the synthesis. The cross-references are already there. The knowledge is compiled once and kept current.
The wiki is a persistent, compounding artifact. The human curates sources and asks questions. The LLM writes and maintains everything.
Three Layers
.raw/ Layer 1 — immutable source documents
wiki/ Layer 2 — LLM-generated knowledge base
CLAUDE.md Layer 3 — schema that tells the LLM how to maintain it
The LLM owns Layer 2 entirely. It creates pages, updates them when new sources arrive, maintains cross-references, and keeps everything consistent. The human reads; the LLM writes.
Operations
Ingest — drop a source into .raw/, tell the LLM to process it. The LLM reads the source, discusses key takeaways, writes a summary page, updates entity and concept pages, and logs the operation. One source typically touches 8-15 wiki pages.
Query — ask a question. The LLM reads the index to find relevant pages, synthesizes an answer with citations. Good answers get filed back into the wiki.
Lint — periodic health check. Find orphan pages, dead links, stale claims, missing cross-references.
Index and Log
index.md — content-oriented. A catalog of all pages with one-line summaries, organized by category. The LLM reads this first on every query to find relevant pages.
log.md — chronological. Append-only record of every ingest, query, and lint pass. Parseable: grep "^## \[" log.md | head -10
Why It Works
The tedious part of maintaining a knowledge base is bookkeeping: updating cross-references, noting when new data contradicts old claims, keeping summaries current. Humans abandon wikis because the maintenance burden grows faster than the value. LLMs don't get bored. The wiki stays maintained because the cost of maintenance is near zero.
At small scale (~100 sources, ~hundreds of pages), the index file is sufficient. No vector database, no embeddings, no infrastructure. Just markdown files.
Comparison to RAG
| Dimension | LLM Wiki | Semantic RAG |
|---|---|---|
| Finding | Reads index, follows links | Similarity search over embeddings |
| Infrastructure | Just markdown files | Embedding model + vector DB |
| Cost | Tokens only | Ongoing compute + storage |
| Maintenance | Run a lint | Re-embed when content changes |
| Scale limit | Hundreds of pages | Millions of documents |
Connections
See Compounding Knowledge for why the pattern produces more value over time. See Hot Cache for the session context optimization. See Andrej Karpathy for the pattern's origin.