LLM Wiki
github.com/nashsu/llm_wiki- Category
- Productivity
- Rank
- No. 767Tools index
Previous survey · No. 781 ·
- Type
- APP
- Use case
- Data, Retrieval & Knowledge
- Interfaces
- Desktop
- GitHub
- 21.0k stars
- Latest release
- v0.6.13
- Date
About
LLM Wiki is a cross-platform desktop application that turns documents into an organized, interlinked personal knowledge base. It uses a two-step chain-of-thought process where an LLM first analyzes the source material and then generates wiki pages with source traceability, and it extracts and captions embedded PDF images using a vision model.
What it does
Drop papers, books, notes and web clips into a folder, and a language model writes and maintains a linked set of Markdown pages about them, which also open as an Obsidian vault. Instead of searching raw files every time you ask something, the model files each new item into existing pages as it arrives, so questions hit notes that were already written. A Rust core behind a React interface handles intake, a link graph, a chat agent, and a local web endpoint that outside coding agents can query.
Why it's ranked here
This is the most complete build of Andrej Karpathy's LLM Wiki pattern, whose original is only an idea file meant to be pasted into an agent. The project ships it as an app under GPL-3.0, with the engineering a long-running tool needs: a local endpoint capped at 120 requests per second, unchanged files skipped by content hash, a persistent intake queue that retries failures up to three times, and parser crashes caught instead of killing the app. Mocked and real-model test suites are wired into the build scripts.
What's good
Every generated page records which original files fed it, so claims trace back to sources. A Read Sources Only mode answers from imported material alone. The web research feature runs a strict relevance check on search results and refuses to finish a write-up that cites none of its collected sources. The bundled agent server reuses the app's own permissions instead of reading folders directly, and page overwrites must be requested explicitly. Graph tools find clusters of related pages and flag weakly connected ones.
Tradeoffs
GPL-3.0 is copyleft, which matters if you plan to ship a modified version. Building from source needs Rust 1.88 or newer plus a heavy stack: a vector database, a PDF engine and office parsers. Each source costs at least two model calls, and the queue runs one item at a time, so large imports are slow and not free. If the research relevance check itself fails, all search results are kept rather than filtered. Generated text is described as English or Chinese only. Read endpoints can be opened without a token.
How to use it well
It suits someone collecting sources on one topic over weeks: a research project, a book, a due-diligence file. Fill in the purpose file first, since the model reads it on every intake and query. Keep token auth on, enable the agent server only if you want Claude Code or Codex to search and write pages, and pin each agent session to one project. For a quick question about a single document, a plain chat upload is cheaper.
Technical notes+
src-tauri/src/api_server.rs binds tiny_http on port 19828 under /api/v1 with RATE_LIMIT_MAX_REQUESTS 120 per one-second window, MAX_IN_FLIGHT_REQUESTS 64, MAX_IN_FLIGHT_CHAT_STREAMS 8, a 2 MiB decoded file cap, and catch_unwind around each handler; streaming agent chat stays token-protected even when unauthenticated reads are allowed. src-tauri/Cargo.toml sets panic = "unwind" and pulls lancedb, pdfium-render, calamine and anydoc. mcp-server/src/index.ts proxies every tool to that API; llm_wiki_write_page needs allow_overwrite. src/lib/deep-research.ts caps sources at 20, runs the relevance judge at temperature 0 with a 45 second timeout, and validateResearchSynthesis rejects output with no [N] citation.
Observed
- License
- GPL-3.0
- Languages
- TypeScript (React frontend) and Rust (Tauri backend)
- Build requirement
- Rust 1.88 or newer declared in the crate manifest
- Interfaces
- Desktop app, local HTTP JSON API, bundled MCP server over stdio
- MCP server requirement
- Node.js 20+ with the desktop app running
- Vector search
- Optional, LanceDB with any OpenAI-compatible embedding endpoint
- Tests
- Vitest suites split into mocked and real-model runs
- Origin
- Implementation of Andrej Karpathy's LLM Wiki pattern
Read from README.md, package.json, LICENSE, llm-wiki.md, src-tauri/Cargo.toml, src-tauri/src/lib.rs, src-tauri/src/api_server.rs, mcp-server/README.md, mcp-server/src/index.ts, src/lib/deep-research.ts.
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.