Your notes. A living wiki. Write in Obsidian — notes, PDFs, web clips. Every input connects to a network that keeps growing.
Lose it, or grow it?
Save an article. Forget where. Search later. Nothing.
Pick a note from your vault. Ideas found, pages built, summary ready. In seconds.
Link two notes by hand. One changes. The other stays old. The link dies.
Links stay alive. When new information contradicts the old, every related page stays current.
A year of notes. Just files. Every project starts over. Same reading. Same conclusions.
A year builds a wiki. Every new question draws on everything you've collected. Your knowledge grows.
One note. A world of connections.
Not features. A new relationship with everything you read.
Pick a note from your vault — `Cmd+P` → Ingest. AI reads, extracts, builds. No folders. No tags. No work.
Ask like you would a friend. Watch it reason, then answer with sources you can open and trust.
Cmd+P → Multi-File Ingest opens a live queue with a folder-tree picker. Select 5 or 50 notes at once — no staging folder needed.
New files? Auto-ingested. Contradictions? Found. Health? Checked. On your terms.
Deep dives on architecture, workflows, and model selection.
Qwen3.5, Qwen3.6, Gemma 4, oMLX — which model runs your wiki well, on what hardware, with what tradeoffs. A practical guide to going fully offline without sacrificing usability.
Pick a PDF from anywhere in your vault. The plugin reads it through your LLM provider, transcribes it to Markdown verbatim, and feeds the result into the standard wiki extraction pipeline.
Four phases that turn a natural-language question into a wiki-grounded answer. Why this pipeline is fundamentally different from vanilla RAG, where it borrows, and where it diverges.
Obsidian is all you need.
In Obsidian: Settings → Community plugins → Browse, search Karpathy LLM Wiki, click Install then Enable.
Open Plugin PageSettings → Karpathy LLM Wiki. Pick your LLM, enter key, test, save.
Cmd+P (or Ctrl+P). Type Ingest to add a single note or Multi-File Ingest to pick many at once. Type Query to ask.
Everything worth knowing, before you start.
When you [[link]] notes manually, you capture connections you already know. The plugin works differently: it reads every note in your vault, extracts the people, concepts, and theories embedded in them, then auto-generates wiki pages that surface links you hadn't noticed — connections between papers you read months apart, themes across folders, insights buried in passing mentions. Your manual links stay intact; the plugin adds an AI-powered layer that finds what you didn't know was there.
Your vault stays exactly as you wrote it. The plugin only creates new wiki pages in its own folders — entities/, concepts/, and similar — under your wiki/ folder. If an auto-generated page is wrong, simply delete it and run ingest again; it will be rebuilt from scratch. Once you mark a page as reviewed, the AI never rewrites it, only appends. Every conversational answer also links back to the source notes it came from, so you can verify any claim with one click.
Obsidian v1.11.0+ (desktop: Windows, macOS, Linux) and an LLM provider API key. Supported: DeepSeek, Gemini, Claude, GPT, Kimi, GLM, MiniMax, OpenRouter, or any custom endpoint. No API key needed for local models via Ollama or LM Studio.
Long-context models recommended — they process your entire wiki in one pass. A single ingest costs $0.05–$0.50. All costs go to the LLM provider; the plugin itself is free and open source.
Yes — and that's exactly how the plugin is built. With Ollama or LM Studio, every byte stays on your machine: notes, extracted entities, wiki pages — never sent, never logged, never touched by any external server. The plugin has zero telemetry, no backend, no storage, and no data collection of any kind; it never connects outbound on its own, even to check for updates. Using a cloud provider? Data is transmitted only during the specific ingest or query you trigger, never cached externally, never shared. This zero-exfiltration architecture means you can confidently use it with research data, client records, personal journals, or any confidential material — regardless of sensitivity. Cloud models offer larger context windows and faster throughput for heavy batch work; local models give you absolute data sovereignty for day-to-day use.
Found a bug? Open an issue on GitHub. Have a question or idea? Start a Discussion on GitHub. You can also find the plugin on the Obsidian Community Plugin page and leave a review.
Yes — since v1.25.0, PDFs are a first-class input. Pick a PDF from anywhere in your vault (`Cmd+P` → Ingest single source). The plugin reads it through your LLM provider's native file input (Anthropic Claude, OpenAI GPT-4o, and Bedrock variants handle PDFs natively), converts it to Markdown via OCR-style verbatim transcription with anti-hallucination markers for figures and equations, and re-enters the regular Markdown ingest pipeline. The converted Markdown is content-hash cached in .obsidian/plugins/karpathywiki/pdf-cache/, so re-ingesting is instant. By default your vault is not modified; enable Write PDF Markdown to Vault in Settings to also write a <basename>.pdf.md sidecar next to the source PDF.
Yes — since v1.24.0, the plugin supports per-task model selection. In Settings → Wiki → Model Scope, switch from Unified (one model for everything) to Per-Task, then pick an independent model for Ingest, Lint, and Query. A common pattern: a cheap fast model for ingest (DeepSeek, MiniMax), a strong model for lint (Claude, GPT-4o), and a long-context model for query. Test Connection probes each configured model sequentially with fail-fast — until every per-task model passes, the connection is considered unhealthy.
Yes — if you pair PDF Ingest with a local model via Ollama or LM Studio, every PDF stays on your machine end-to-end. The recommended setup on Apple Silicon is oMLX + the Markitdown backend + Baidu's open-sourced Unlimited-OCR model (3B total / 0.5B active), connected via a Custom OpenAI-Compatible provider with Force PDF Support enabled. PDFs never leave your vault; conversion runs locally; the resulting Markdown hits the same entity-extraction pipeline, also on-device.
Not an island. Every tool you love, now better.
See links
Wiki gives your graph meaning. Hub pages, clusters, orphans. Exploration, not decoration.
Save articles
Save the clip anywhere in your vault. Run `Cmd+P` → Ingest from folder. AI extracts, links, updates. One clip becomes 10+ wiki pages.
Query data
AI adds structure. Tags, dates, categories. Empty tables become living dashboards.
Track versions
Watch your knowledge evolve. Every commit tells the story of how you grew.
Make slides
Ask: "Summarize my research for a 10-minute talk." Wiki builds slides from your knowledge.
Visual canvas
AI builds concept maps, timelines, decision trees from your knowledge. No more blank canvas.
Twelve providers, cloud or local. Switch anytime — lock into nothing.
Twelve providers, one dropdown apart. Outgrow one, switch to the next — your wiki never notices.
No backend. No tracking. Nothing stored. Your notes only travel to the AI you choose — or stay fully offline with a local model.
The plugin is free. A full ingest runs you $0.05–$0.50 on the LLM side — pick a value model and you'll barely notice it.
Cloud or local — use the one you trust.
Long-context models read your whole wiki in a single pass — the complete picture, not fragmented snippets. Ollama and LM Studio run fully on-device — no upload, no API key.
