llm-wiki · GitHub

The core thesis contrasts the LLM Wiki pattern with traditional RAG systems. Instead of retrieving and re-synthesizing from raw documents on every query, the LLM compiles knowledge once into a persistent wiki and keeps it current. This means cross-references, contradictions, and synthesis are pre-computed rather than re-derived each time, making the knowledge base a compounding asset.

The system consists of three distinct layers — (1) Raw Sources (immutable documents the LLM reads but never modifies), (2) The Wiki (LLM-generated and maintained markdown files with summaries, entity pages, and cross-references), and (3) The Schema (a configuration document like CLAUDE.md that tells the LLM how to structure and maintain the wiki). This separation of concerns ensures source integrity while allowing the wiki to evolve.

Ingest processes new sources by reading them, creating summary pages, and updating existing wiki pages (a single source might touch 10-15 pages). Query searches the wiki and synthesizes answers that can themselves be filed back as new wiki pages. Lint is a periodic health-check that finds contradictions, orphan pages, stale claims, and missing cross-references.

The fundamental reason the pattern works is that LLMs eliminate the maintenance burden that kills human-maintained wikis. Updating cross-references, keeping summaries current, noting contradictions, and maintaining consistency across dozens of pages — this bookkeeping grows faster than value for humans but costs near-zero for LLMs. The human's role is curating sources, directing analysis, and asking good questions.