llm_wiki

This repository is a minimal implementation of the pattern described in llm-wiki.md. It is designed to work with a coding agent and an Obsidian vault without requiring a database, embeddings, or a hosted service.

Start here

• Put a document in raw/inbox/. Keep the original filename where practical.
• Ask your agent: Ingest raw/inbox/ into the wiki.
• Review the source note and the changes to topic/entity pages.
• Move accepted sources from raw/inbox/ to raw/library/ if you want the

inbox to represent unprocessed work. Moving is a human action; agents treat everything under raw/ as read-only.

• Ask questions normally. To preserve an answer, say: `File this analysis in

the wiki.`

• Run python3 scripts/wiki_lint.py periodically.

Browser interface

Install dependencies once, then start the local application:

Open . The interface is read-only and updates when a Markdown file under wiki/ changes. It includes full-text search, backlinks, raw-source links, and a relationship graph. Press Ctrl/Cmd+K to focus search.

The server binds to loopback and rejects non-local Host headers. It is not an authenticated multi-user service; do not publish it through a reverse proxy or tunnel without adding authentication at that boundary. Raw-file requests are resolved to their real path and rejected if a symlink escapes raw/.