AI Wiki - A personal knowledge wiki where an LLM builds up papers, news and repos (논문, 뉴스, 저장소를 LLM이 쌓아올리는 개인 지식 위키)
Translated by AIView original
Screenshots
Screenshot 1 of 4
This is my implementation of the LLM Wiki idea that Andrej Karpathy proposed. Instead of searching with RAG and answering from scratch every time, the LLM writes wiki pages at the moment you add material, and questions are answered using only that wiki. I set four rules in CLAUDE.md to block hallucinations: no web search, the wiki is the only source of truth, if it's not enough then re-read the original and update, and if there's no material then say there's no material. Those four rules are the core.
The material is organized in three layers. raw/ keeps PDFs, repo snapshots and original articles as they are, and sources/ stores summaries the LLM wrote. wiki/ sorts the organized pages into preset categories (llms, agents, physical-ai, etc.).
The whole process, from collecting material and summarizing to writing the wiki and linting terms, is automated with Claude Code skills and hooks.
I also added a web deployment so you can read it from anywhere. I built the markdown into static HTML with Node.js, added full-text search with Pagefind and a graph of page relationships, and it auto-deploys to github.io through GitHub Actions. There are about 290 pages of knowledge uploaded so far, and I use it for personal study.
How it was made
Start: I read papers and still didn't retain anything
I keep up with AI papers, news and repos, but after reading I could never remember where things were. I tried organizing them in Notion too, but writing the documents took too much time, so I gave up.
Then I came across the LLM Wiki concept that Andrej Karpathy proposed. The key idea is that the LLM writes wiki pages at the moment you add material, not at search time, and questions are answered using only that wiki. That's it, I thought, so I decided to build it myself with Claude Code.
Design: three layers and four rules
I borrowed from other repo implementations and split the folders into three layers.
raw/ : PDFs, repo snapshots and article originals, kept untouched sources/ : Structured Korean summaries written by the LLM (summary, contributions, methods, results, limitations, related work, terms) wiki/ : Refined pages linked with [[wikilinks]]
I unified the filename stem across all three layers (something like vaswani-2017-attention-is-all-you-need), so you can always trace back from the wiki to the original.
And I set four rules in CLAUDE.md: no web search, the wiki is the only source of truth, if it's not enough re-read the original and update the wiki, and if there's no material then say so.
I sorted the content by my four research interests, not by topic. RAG infrastructure goes under database, robot learning under physical-ai, benchmarks under evaluations, and so on. I based it on "where would future me want to look for this?"
Implementation: automating the pipeline with Claude Code skills and hooks
I made skills for each material type. Skills like ingest-paper and ingest-article handle text extraction (pypdf, pymupdf), writing sources, writing the wiki and updating index.md all in one go.
The part I cared about most was figure handling. I extract every figure from a paper, but the LLM proposes candidates in a table, and I copy only the ones I've checked into wiki/assets/ and put them in the wiki. Adding everything automatically makes the wiki messy, and adding nothing makes it hard to understand.
Quality control I left to hooks and linting. After saving, wiki-lint-reminder.sh runs, lint_terms.py checks that terms match the glossary (wiki/overviews/glossary-*.md) and notation, and lint_style.py checks that the style rules are followed. This is where I fix the problem of notation wobbling when English technical terms are mixed into Korean text.
- Built with
Comments 0
No comments yet
Be the first to comment.