Karpathy LLM Wiki is an 📝Obsidian community plugin that implements 📝Andrej Karpathy's LLM wiki pattern inside a vault, compiling notes and PDFs into interlinked entity and concept pages without using embeddings.
The plugin turns a workflow into a product. Where 📝Karpathy's Flag in the Ground — LLM Knowledge Bases described a folder structure driven by a coding agent and a schema file, the plugin ships the same three operations — ingest, query, and lint — as buttons inside the note-taking app. It extracts entities and concepts into standalone pages, detects duplicates across languages, flags contradictions between sources, and runs periodic health scans over the resulting graph. It is published under Apache 2.0, runs entirely on the user's machine, and never modifies the original vault notes.
Key Facts
- Category: Obsidian community plugin for LLM-compiled knowledge bases
- Repository: green-dalii/obsidian-llm-wiki; community directory ID
karpathywiki - License: Apache 2.0, free, with no backend and no analytics
- Retrieval: Zero-embedding by design — uses the vault's existing wikilink graph
- Providers: 12+, including Anthropic, OpenAI, Google Gemini, DeepSeek, and local models via Ollama and LM Studio
- Localization: 10 interface languages, with wiki output language selected independently
How It Works
Compilation. Sources dropped into the vault are read by the configured model and compiled into entity and concept pages with bidirectional references. Aliases are mandatory on generated pages, which is what lets the plugin catch the same entity appearing under different names or in different languages. Merge detection flags contradictions between sources rather than silently overwriting.
Graph retrieval without vectors. Queries run through a five-stage cascade: lexical match, LLM-generated keywords, substring scan, semantic fallback, and finally Personalized PageRank expansion across the wikilink graph, implemented as Monte Carlo random walks. The design choice is deliberate — hand-curated links are treated as higher-quality edges than computed similarity, which removes any embedding infrastructure from the stack. The project publishes a benchmark of 27.1% at rank 5 against a 24.1% pure-kNN baseline.
Lint and repair. A health scan surfaces duplicates, dead links, empty pages, and orphans, with a one-click repair that applies fixes in dependency order. Operation history records what the plugin changed. This is the direct implementation of the third operation in Karpathy's description, the one most often skipped in hand-rolled versions.
Document ingest. Version 1.25.0 added PDF handling with native support for Anthropic, OpenAI, and Bedrock, plus optional cloud or local OCR paths for documents that need it.
Why It Matters
The plugin is the clearest evidence that the gap Karpathy named — "an incredible new product instead of a hacky collection of scripts" — was real and is being filled from inside the existing tools rather than by a new platform. It arrived free, local, and feature-complete against his description within months of the post.
Its architecture also marks the boundary of what a plugin can do. Everything runs on one machine for one user: no permission model over what a model may read, no access from a phone or a cloud agent, and no shared or collaborative layer. Those are not omissions in the plugin; they are the ceiling of building inside a local-first note-taking app.
FAQ
What is the Karpathy LLM Wiki plugin?
An Obsidian plugin that compiles notes and PDFs into an interlinked wiki of entity and concept pages, following the LLM knowledge base pattern Andrej Karpathy described in April 2026.
Does it use embeddings or a vector database?
No. It is zero-embedding by design, ranking results by expanding across the vault's existing wikilink graph rather than by computing vector similarity.
Does it modify existing notes?
No. Generated wiki pages are written alongside the vault's original notes, which the plugin leaves untouched.
Which models can it use?
More than twelve providers, including hosted APIs from Anthropic, OpenAI, Google, and DeepSeek, and local models served through Ollama or LM Studio.
Related
- 📝Karpathy's Flag in the Ground — LLM Knowledge Bases — the original architecture this plugin implements
- 📝Obsidian — the application the plugin runs inside
- 📝Obsidian MCP — the alternative route to AI-in-a-vault, via external agents rather than an in-app plugin
- 📝The Karpathy Weathervane Pattern — why a single post from this source reorganizes a field within weeks