wiki-wonka
LLM Wiki
A persistent, LLM-maintained knowledge base inspired by Karpathy's LLM Wiki pattern.
🧭 Overview
wiki-wonka is an agent-driven system for building and maintaining a structured, evolving wiki of knowledge. Instead of relying on stateless retrieval (RAG) for every query, wiki-wonka compiles and curates knowledge into a persistent, interlinked set of Markdown pages. The LLM agent handles all the summarizing, cross-referencing, and bookkeeping—so your knowledge base compounds and improves over time.
🛠️ How it works
- Raw sources: Immutable documents (papers, articles, data) are added to the
raw/ folder. The LLM never edits these files.
- Wiki: The agent generates and maintains Markdown pages in
wiki/, including summaries, entity pages, concepts, and an evolving synthesis. All cross-references and updates are handled automatically.
- Schema: The structure and conventions are defined in a schema file (see
wiki/SCHEMA.md). This ensures consistency and enables the LLM to act as a disciplined maintainer, not just a chatbot.
✨ Features
- Persistent knowledge base — wiki pages compound over time; no rediscovery from scratch on every query
- Structured ingestion — two-phase flow (analysis + confirmation) before any file is written
- Cross-referencing — wikilinks (
[[slug]]) connect sources, entities, and concepts automatically
- Query with citations — answers always sourced from wiki pages, never from model memory alone
- Lint & health checks — detects contradictions, orphan pages, outdated claims, and missing cross-references
- Append-only log — full audit trail of every ingest, query, and lint operation in
wiki/log.md
- Callout system —
[!contradiction], [!gap], [!outdated], [!deprecated] flag issues for human review
- Language support — all generated prose follows the locale set in
wiki/config.md (e.g. pt-BR, en-US)
- Git auto-commit — hook commits wiki page writes automatically, keeping history clean
- Immutable raw sources —
raw/ is write-protected; the agent can never modify original documents
- Schema-driven — conventions defined in
wiki/SCHEMA.md keep every page consistent and predictable
⚡ Main operations
- Ingest: Add a new source to
raw/ and instruct the agent to process it. The agent reads, summarizes, updates relevant pages, and logs the operation.
- Query: Ask questions against the wiki. The agent synthesizes answers from existing pages, always citing sources. New insights can be filed back into the wiki.
- Lint: Periodically check the wiki for contradictions, stale claims, orphan pages, and missing cross-references. The agent suggests fixes and keeps the knowledge base healthy.
🌐 Language support
wiki-wonka generates all wiki content in the language configured in wiki/config.md:
---
language: pt-BR # e.g. en-US, pt-BR, es
---
The orchestrator reads this file at startup and enforces the language across every operation — ingest summaries, query answers, concept definitions, lint reports, and free-conversation responses. To switch languages, edit the file and continue working; existing pages are not rewritten automatically.
What always stays in English regardless of the setting:
- Frontmatter field names (
title, slug, type, tags, …)
- File slugs (
attention-mechanism.md)
- Callout types (
[!contradiction], [!gap], [!outdated], [!deprecated])
🗂️ Indexing and logging
wiki/index.md: Catalog of all wiki pages, organized by category, with summaries and links.wiki/log.md: Chronological log of all ingests, queries, and maintenance actions.
🤔 Why this approach?
Traditional RAG systems force the LLM to rediscover knowledge from scratch on every query. By maintaining a persistent, evolving wiki, wiki-wonka enables deeper synthesis, better cross-referencing, and a continuously improving knowledge base. The LLM does the grunt work; you curate sources and guide the process.
💡 Usage
Before interacting, start the agent server in claude code:
# 1. navigate to the project root
cd wiki-wonka
# 2. start the agent server with the current directory as the plugin source
claude --plugin-dir .
[!NOTE]
You can then interact with the agent wiki-wonka through any interface that supports the plugin, such as a chat UI or Copilot.
The user interacts directly in natural language—no special commands to memorize. The orchestrator interprets intent and routes to the right skill.
Here are the main interaction patterns:
Ingest
The user mentions a source—file path, pasted URL, or direct content:
"process this article: raw/attention-is-all-you-need.md"
"I just saved a PDF in raw/, it's called transformers-survey.pdf"