CONCEPTS

Llm Wiki Pattern

--- title: LLM Wiki Pattern created: 2026-05-25



title: LLM Wiki Pattern
created: 2026-05-25
updated: 2026-05-25
type: concept
tags: [wiki-pattern, pkm, knowledge-base]
sources: [raw/articles/llm-wiki-pattern-introduction.md]
confidence: high
contested: false
contradictions: []

The LLM Wiki Pattern

The LLM Wiki Pattern is a novel methodology for building and maintaining personal or team pkm|knowledge bases using LLMs. Developed as a compounding, persistent alternative to standard rag, it shifts the effort from "retrieving and synthesizing on every query" to "compiling knowledge once and keeping it current."

Core Idea: Compounding vs. Rediscovering

In traditional RAG or doc-chat interfaces (e.g., NotebookLM, ChatGPT uploads), the LLM starts from scratch for every question. It must find relevant fragments, resolve contradictions, and build a synthesis in real time, forgetting everything once the session ends.

The LLM Wiki Pattern introduces a persistent, interlinked wiki (using markdown files) that sits between the raw sources and the user. The LLM acts as the curator, librarian, and programmer of this wiki:

  • Incremental Integration: When a new source is added, the LLM reads it, updates relevant entity/concept pages, notes contradictions, and links information together.

  • Compounding Artifact: The knowledge is pre-compiled. Cross-references are resolved, and contradictions are flagged before you ask a question.

  • Obsidian Integration: The markdown-based repository functions natively as an Obsidian vault, allowing the human to browse pages, click links, and view the graph network in real time.

Division of Labor


  • The Human: Sourcing raw materials, exploring the network, directing analysis, and guiding the synthesis.

  • The LLM Agent: Heavy lifting, summarizing, bookkeeping, link validation, formatting, and file indexing.

Architecture: The Three Layers

1. Raw Sources (Layer 1): Immutable files (articles, transcripts, PDFs, images) that serve as the ground truth. The LLM reads these but never edits them.
2. The Wiki (Layer 2): Directory of LLM-generated markdown files, categorised into entities/, concepts/, comparisons/, and queries/.
3. The Schema (Layer 3): Standard rules, file-naming conventions, and metadata structures (e.g., SCHEMA.md) that make the LLM a disciplined wiki maintainer.

Core Operations

  • Ingest: Reading a raw source (including audio/video sources through media-ingestion), updating affected entity/concept pages (often touching 10-15 files), and logging the changes.

  • Query: Answering questions based on the pre-compiled wiki rather than searching raw chunks, and filing high-value syntheses back into the wiki.

  • Lint: Auditing the wiki to find orphan pages, dead links, missing tags, stale claims, and unresolved contradictions.

Navigation & Tracking


  • index.md: Content-oriented, alphabetical list of all pages and one-line summaries.

  • log.md: Chronological record of all actions (e.g., ## [2026-05-25] ingest | Article Title).

Related Knowledge