I also adopted an index‑first retrieval strategy: the agent scans a Markdown index of page titles and short descriptions, picks the most relevant pages, and follows the links for deeper information, eliminating the need for a vector store. The ingestion pipeline follows five deterministic steps – copy the original document into raw/, create a source page with key insights, use an LLM to extract entities and concepts, merge recurring concepts into dedicated Markdown pages, and finally rebuild the indexes with a Python script. Incremental updates only recompute pages that depend on a changed raw source, keeping rebuild times low.
My LLM wiki uses Markdown indexes containing page titles and descriptions. The agent reads the index, chooses relevant pages, and follows links when it needs more information.
A concept only gets its own page when it appears in at least 2 different sources.
Copy the original document into raw/. Create a source page with the key insights. Use an LLM to extract entities and concepts. Merge recurring concepts into dedicated Markdown pages. Rebuild the indexes using a Python script.
It’s basically a lightweight version of GraphRAG built entirely with Markdown files.
One wiki contains: • 196 raw sources • 7 GitHub repositories • 399 knowledge nodes
In my upcoming Decoding AI Magazine, I show you how to build an LLM wiki from first principles.
Pele is a new rendering library for diagrams/charts, compatible with Mermaid syntax, 100‑500× faster than mermaid.js – 6× smaller bundle – No dependencies – Easy to theme – Responsive to device width.
In October 2026 I built an LLM‑wiki as a lightweight, Markdown‑only knowledge graph to give AI agents persistent memory without blowing past context windows.
The wiki contains 399 knowledge nodes.
It aggregates 196 raw sources.
The collection pulls from 7 GitHub repositories.
A concept only gets its own page when it appears in at least 2 different sources.
I chose pure Markdown for its editability and used an incremental rebuild pipeline so only pages linked to changed sources are regenerated, keeping processing time low.