A few days ago, Andrej Karpathy’s post on "LLM Knowledge Bases" went viral. He proposed a shift from manipulating code to manipulating knowledge-using LLMs to incrementally compile raw data into a structured, interlinked graph of markdown files.
I loved the idea and started testing it out. It worked incredibly well, and I decided this was how I wanted to store all my research moving forward. But the friction was killing me. My primary device is my phone, and every time I found a great article or paper, I had to wait until I was at my laptop, copy the link over, and run a mess of scripts just to ingest one thing.
I wanted the "Knowledge wiki" in my pocket. 🎒
I’m not a TypeScript developer, but I decided to "vibecode" the entire solution into a native app using Tauri v2 and LangGraph.js. After a lot of back-and-forth debugging and iteration, I’ve released LLM Wiki.
How it works with different sources:
The app is built to be a universal "knowledge funnel." I’ve integrated specialized extractors for different media: * PDFs: It uses a local worker to parse academic papers and reports directly on-device. * Web Articles: I’ve integrated Mozilla’s Readability engine to strip the "noise" from URLs, giving the LLM clean markdown to analyze. * YouTube: It fetches transcripts directly from the URL. You can literally shared a 40-minute deep-dive video from the YouTube app into LLM Wiki, and it will automatically document the key concepts and entities into your graph while you're still watching.
The "Agentic" Core:
Under the hood, it’s powered by two main LangGraph agents. The Ingest Agent handles the heavy lifting of planning which pages to create or update to avoid duplication. The Lint Agent is your automated editor—it scans for broken links, "orphan" pages that aren't linked to anything, and factual contradictions between different sources, suggesting fixes for you to approve.
Check it out (Open Source):
The app is fully open-source and brings-your-own-key (OpenAI, Anthropic, Google, or any custom endpoint). Since I vibecoded this without prior TS experience, there will definitely be some bugs, but it’s been incredibly stable for my own use cases.
GitHub (APK and EXE in the Releases): https://github.com/Kellysmoky123/LlmWiki
If you find any issues or want to help refine the agents, please open an issue or a PR. I'd love to see where we can take this "compiled knowledge" idea!
[link] [comments]
