llm-wiki-expert

Build AI experts backed by evolving wiki knowledge bases — not static RAG.

Inspired by Karpathy's LLM Wiki idea. Instead of retrieval-augmented generation that re-derives knowledge from scratch every query, this system lets LLMs continuously build and maintain a structured wiki between you and your sources.

Karpathy's 4 Principles (implemented here)

• Explicit — knowledge is visible, navigable, inspectable
• Yours — data stays local, not in any AI provider's system
• File over app — markdown files work with Obsidian, Cursor, Claude, anything
• BYOAI — plug in any OpenAI-compatible model

Features

• 📥 Ingest anything — paste text, import URLs, drop AI conversations
• 🔗 URL import — fetch any page, LLM extracts knowledge automatically
• 💬 Conversation import — compile AI chats into permanent wiki knowledge
• 🕸️ Knowledge graph — visual map of concepts and connections
• 📤 Obsidian export — markdown with frontmatter + [[wiki links]]
• 🔍 Gap detection — LLM identifies what the expert doesn't know yet
• 🤖 Chat with your expert — answers grounded in wiki, not hallucination

Quick start

Open create an expert, start ingesting.

Architecture

The LLM incrementally updates concepts — no overwriting, always compounding.

Ingestion strategies

For long documents, two strategies auto-selected: