llm-wiki-flywheel

Language / 语言： English · 简体中文

Compile, don't retrieve. Drop a source in. Claude does the rest: extract entities, build wiki pages, inject wikilinks, track trust, flag contradictions. Markdown-first; optional hybrid retrieval. Pure markdown you own, browsable in Obsidian.

Inspired by Karpathy's LLM Knowledge Bases, then fully automated. Works natively inside Claude Code via 28 MCP tools, no API key required. Also runs on any local AI CLI tool (Ollama, Gemini CLI, OpenCode, Codex CLI, and more) via KB_LLM_BACKEND.

🎯 What you get

• 🧠 Structure first, optional vectors. Entities, concepts, wikilinks form a real graph; hybrid BM25 + vector search is opt-in for recall.
• ⚡ Incremental by default. SHA-256 change detection; only new/changed sources reprocessed.
• 🔗 Retroactive linking. Ingest a new topic → existing pages auto-gain [[wikilinks]] to it.
• 🧪 Self-healing. Bayesian trust scoring, contradiction detection, staleness flags, dead-link lint.
• 🦉 Obsidian-native. Open wiki/ as a vault for a free graph view, backlinks, and hover preview.
• 🔌 MCP-first. 28 tools in Claude Code. Talk to your wiki: "ingest this", "what do we know about X?"
• 📤 Publishable. One command emits /llms.txt, /llms-full.txt, /graph.jsonld, sitemap, and per-page siblings: the Karpathy Tier-1 machine-consumable stack.

Why Not RAG?

RAG retrieves chunks. This system understands structure.