liveBook · Manning

LLM wiki

As explained in Retrieval Augmented Generation, The Foundational Ideas

The LLM Wiki pattern compiles a corpus into a persistent, interlinked Markdown knowledge base that an agent reads to answer questions.

The LLM Wiki pattern compiles a corpus once into a persistent, interlinked Markdown knowledge base. An agent then answers questions by reading relevant wiki pages instead of retrieving fragments from the raw documents. section 10.7.3

The pattern shifts work from query time to corpus compilation:

  • An agent reads the corpus, creates structured pages, and links related topics. section 10.7.3
  • Query-time behavior becomes selective reading of human-auditable pages rather than vector search over raw documents. section 10.7.3
  • Clean structural boundaries make the approach more viable; unstructured prose provides less hierarchy to surface. section 10.7.3

Suitable corpus

Research papers, manuals, and contracts can supply stable sections and topics for linked pages. The resulting wiki can be improved incrementally, but drift from the sources may be difficult to detect because the prose appears authoritative. section 10.7.3

Treat the pattern as experimental rather than automatically superior: published implementations are young, and accepted evaluations against retrieval baselines are not yet available. section 10.7.3

How Retrieval Augmented Generation, The Foundational Ideas covers it

How deep it goes
Explains the architecture, applicability boundary, and drift risk, but stops short of implementation detail or validated evaluation.
Examples
Uses research papers, manuals, and contracts as corpora with clean structural boundaries.
What it assumes
Assumes familiarity with agents, retrieval, vector search, and corpus structure while treating source auditing as an unresolved concern.
What you can do after
Evaluate whether persistent wiki compilation fits a structured corpus without assuming it outperforms retrieval.
How it connects
Connects compilation with RAPTOR, and contrasts it with HIRO, SiReRAG, CRP-RAG, and vector search.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Retrieval Augmented Generation, The Foundational Ideas. The section links point to the parts of the book it draws on.