AutoAtlas

Andrej Karpathy described an LLM-powered knowledge base workflow and said there was room for an incredible new product. AutoAtlas is my attempt at that product: a Python tool that watches source folders and uses Claude to compile them into a structured Markdown wiki.

Why a compiled wiki, not RAG

Retrieval-augmented generation chunks documents, embeds them and retrieves fragments at query time. Chunking destroys context, and nothing accumulates.

A compiled wiki has the LLM read the raw documents and write a cross-linked, human-readable knowledge base in plain Markdown. It grows with every document ingested, and questions are answered by navigating it like a researcher would. No vector database.

How AutoAtlas works

Ingest: it watches folders and extracts text from PDFs, Word documents, Markdown, text files and DWG drawings, skipping unchanged files. Compile: Claude writes new and updated pages and cross-references, committed to git. Query: a chat interface returns cited answers from the wiki. Lint: periodic health checks find broken links, stale claims and contradictions.

It also runs on local models through Ollama, since the documents it is built for probably should not leave your network.

What comes next

AutoAtlas is a proof of concept, validated end to end. The obvious next steps are hosted deployment, team authentication, and connectors for SharePoint and Google Drive.