Ideas

The LLM wiki pattern

Instead of re-discovering knowledge on every question, a language model keeps a wiki of linked pages current, so knowledge builds up over time.

The idea

Many AI systems look things up from scratch on every question. They search raw documents, answer, and forget. The LLM wiki pattern, taken from a widely shared public idea file, works differently.

A language model reads each new source once. It writes what it learned into a lasting wiki, updates the pages it touches, and flags contradictions. When a question comes, the links and the summary are already there. The person picks sources and asks questions. The model does the bookkeeping.

Three layers

• Raw sources, which never change.
• The wiki, written and kept by the model.
• The schema, the rules for the wiki, shaped by both the person and the model over time.

Three operations

• Ingest. Read a new source and update the pages it affects.
• Query. Search, read and answer with citations. A good answer is filed back into the wiki.
• Lint. Look for contradictions, stale claims, orphan pages and missing pages.

Why it works

Knowledge compounds. Each source makes the wiki a little better, and the work of connecting ideas is done once instead of on every question.

See Wiki for the knowledge base Kinesis keeps.