AutoWiki
Your LLM Compiles a Knowledge Base. You Just Read and Ask.
Implementing Karpathy's LLM Knowledge Base vision — raw sources in, Obsidian wiki out.
By AlphaLab-USTC
English · 中文
The Idea
Good survey papers don't just list references — they identify milestones, trace how ideas evolved, and organize the field into a coherent structure. AutoWiki does the same thing, automatically.
Drop papers in. The LLM identifies milestone nodes (conceptual breakthroughs), clusters papers around them, builds a hierarchical file tree that mirrors the field's structure, and writes deep analysis with temporal links between sources. You get a survey-quality knowledge graph maintained in Obsidian — not a folder of disconnected notes.
You drop PDFs LLM compiles You browse in Obsidian
───────────── ────────────────────────── ──────────────────────────────────
raw/new/ → Identify milestones kb/topics/ (milestone nodes)
Cluster papers kb/sources/ (deep analysis pages)
Trace temporal evolution kb/journal/ (cognitive timeline)
Write cross-linked wiki index.md (survey-style tree)
🚀 Quick Start
1. Clone & Install
git clone https://github.com/AlphaLab-USTC/AutoWiki-skill.git
pip install PyMuPDF
2. Add as Skill
In Claude Code or OpenClaw, just ask:
> "Add /absolute/path/to/AutoWiki-skill/skills to my global skillSources"
3. Initialize Your Wiki
cd your-wiki-project
mkdir -p raw/new raw/compiled kb/{sources,topics,journal} output
touch kb/index.md kb/log.md
4. Start Using
Open the project root as an Obsidian vault, then in Claude Code / OpenClaw:
> "Ingest the paper in raw/new/"
> "How does X compare across the Y papers?"
> "Lint the wiki"
Prerequisites: Claude Code or OpenClaw + Obsidian + Python 3.12+
How It Works
raw/new/ ──→ LLM (Claude Code) ──→ kb/ ──→ Obsidian
you drop reads, analyzes, sources/ you browse
PDFs here links, writes topics/
journal/
| Layer | Role | Who owns it |
|---|
raw/ | Source archive (PDFs + figures) | You drop files; agent organizes |
kb/ | Living wiki (markdown + wikilinks) | Agent writes & maintains everything |
Three operations: Ingest (PDF → analysis page), Query (ask questions → synthesize → write back), Lint (25+ auto-checks).
What Makes It Different
| |
|---|
| 🧠 Deep Analysis | Not summaries. CRGP factors from the author's Introduction. Critical analysis with prior/update contrastive structure. Anti-patterns prevent generic filler. |
| 🔗 Temporal Graphs | Every source positioned in the field's timeline. Evolutionary chains, cross-domain links, temporal tensions — all automatic. |
| 🏠 Self-Healing | Three-tier autonomy: Silent (fix links) → Notify (record insights) → Confirm (restructure). 25+ lint checks. |
| 🎯 Smart Classification | 3-question fitness check before every topic assignment. Auto-scaling: <5 inline, ≥5 split, >8 sub-cluster. |
📂 Showcase: 80-Paper Wiki
We built a real wiki on Agent Self-Evolution — 80 papers, 13 milestones, 2 hours. Here's what the LLM compiled (compiling 80 papers in Claude Code with Opus 4.6 cost ~150 🔪 of quota):
