Productivity · Framework
Andrej Karpathy's LLM Knowledge Base Workflow
A workflow for using LLMs to compile and maintain personal wikis
This framework outlines a concrete system for building an autonomous, LLM-maintained personal wiki in markdown using Obsidian as the interface. Rather than relying on complex RAG pipelines, an LLM incrementally processes raw source documents, generates cross-linked concept pages, answers complex research queries, and lints the wiki for gaps. It treats knowledge management like a compiled codebase where the model handles all the editing and structuring.
The framework
Workflow: (1) Data ingest — index source documents (articles, papers, repos, datasets, images) into a raw/ directory; use an LLM to incrementally "compile" a wiki of .md files: summaries of all raw/ data, backlinks, categorization into concepts, articles per concept, cross-links. Use Obsidian Web Clipper to convert web articles to .md and a hotkey to download related images locally so the LLM can reference them. (2) IDE — Obsidian as the frontend for raw data, compiled wiki, and derived visualizations; the LLM writes and maintains all wiki data, the human rarely edits directly; plugins like Marp render data as slides. (3) Q&A — once the wiki is big enough (his research wiki: ~100 articles, ~400K words), ask the LLM agent complex questions; no fancy RAG needed because the LLM auto-maintains index files and brief summaries and reads related data easily at this scale. (4) Output — answers rendered as markdown files, Marp slide shows, or matplotlib images, viewed in Obsidian; outputs are often "filed" back into the wiki so explorations "add up". (5) Linting — run LLM health checks to find inconsistent data, impute missing data via web search, and surface connections as new article candidates; LLMs also suggest further questions. (6) Extra tools — e.g. a vibe-coded naive search engine over the wiki, used directly in a web UI or handed to the LLM via CLI as a tool. (7) Further explorations — synthetic data generation + finetuning so the LLM "knows" the data (truncated).
What it's for
- Compile unstructured research papers and articles into an interconnected markdown wiki
- Run autonomous agent Q&A over personal notes without complex vector databases
- Audit a personal knowledge base for inconsistencies, missing data, and new connections
- Generate presentation slides and visual summaries directly from compiled personal notes
Key points
- Raw data ingestion using tools like Obsidian Web Clipper and local asset downloads
- Automated wiki compilation where the LLM maintains summaries, backlinks, and categorization
- Obsidian-based IDE interface where the human reads and the LLM writes
- Index-driven Q&A bypassing complex RAG by having the agent read structured index summaries
- Recursive output generation that saves derived answers, slides, and plots back into the wiki
- Automated wiki linting to impute missing information and discover new concept candidates
- Custom tooling support via CLI search tools or local web interfaces
Pricing
Free
Details
- Type
- Framework
- Category
- Productivity
- Topic
- Knowledge Management
- Open source
- No
- Platforms
- other
- For
- Researchers, developers, and knowledge workers wanting an LLM-first notes architecture.
- Alternative to
- Mem.ai, Notion AI
- Tags
- pkm, obsidian, llm, agents, workflow, markdown
Why it was saved
- @karpathy: A concrete, opinionated recipe: ingest sources, let the LLM compile a wiki, use Obsidian as the frontend, run agent Q&A over it, render outputs back into the wiki, and lint for integrity.
Compare
Related
Connected workspace for notes, docs, databases, and AI automation.
Ad-free platform for saving, organizing, and connecting ideas into channels
Local-first IDE and framework for building typed, agentic knowledge graphs
Architecture framework for building an LLM-maintained markdown wiki that compounds knowledge
A design pattern for building LLM-maintained personal knowledge bases in markdown
Turn Twitter bookmarks into a searchable AI knowledge base with citations
Updated