Every imported file folds into a wiki the model maintains
LLM Wiki is a Tauri desktop app, React on top of a Rust backend, that hands an LLM the job of keeping a wiki built from your documents. The model summarizes each source, spreads what it found across entity and concept pages joined by [[wikilinks]], and updates the pages that already exist. Chat answers draw on keyword search over the wiki and the raw sources, optional LanceDB vector search, and a two-hop walk along the link graph. Each project writes its own .obsidian/ config and opens as an Obsidian vault.
Source-traced pages and a four-signal relevance graph
- Two-call ingest: the first LLM call analyzes the source (entities, arguments, contradictions with the wiki), the second writes the pages. Every generated page lists the files it came from in a
sources[]frontmatter field. - Cascade delete: removing a source deletes its summary page and strips it from the
sources[]of shared pages, which stay in place. Dead index entries and dead wikilinks go with it. - Weighted graph: direct links count x3, shared sources x4, Adamic-Adar x1.5, same page type x1. Louvain clustering groups pages into communities and flags any whose cohesion drops below 0.15.
- Research aimed at graph gaps: isolated pages, loosely linked communities and bridge nodes get a Deep Research button. The topic can be edited before the run, the search goes through Tavily, SerpApi or SearXNG, and the findings land back in the wiki.
- Formats: PDF (bundled Rust extractor or MinerU), DOCX, PPTX, XLSX, EPUB, MOBI, Org mode, and pages clipped with the Chrome extension. A SHA256 hash keeps unchanged files from going back to the model.
- Local API and MCP server: a JSON API on
127.0.0.1:19828and an MCP server expose hybrid search, file reads, the graph and the agent chat. A skill from a separate repo plugs Claude Code or Codex into it, read-only by default.
Two LLM calls per source, Apple Silicon Macs only
Nothing gets built until a model is configured. The choices are OpenAI, Anthropic, Google, a custom endpoint, or Ollama to keep inference local. Every source costs two calls, and the ingest queue runs them one at a time, retrying a failed task up to three times.
Intel Macs lost their installer after 0.2.0. Every macOS release since then ships for aarch64 only, while Windows gets x64 packages and Linux gets .deb, .rpm and AppImage builds for amd64 and arm64. Building from source takes Node.js 20+, Rust 1.88+ and protoc.
The README lists fewer interface languages than the app has. It names English and Chinese; the translation files in src/i18n/ add Italian and Russian. The language the wiki is written in is a separate setting, picked at project creation from a list of 23.
The README pitches a personal knowledge base. The API listens on 127.0.0.1 only, a project moves to another machine through a ZIP export, and the Chrome extension is loaded unpacked in developer mode.
The project sets itself against classic RAG. It implements Andrej Karpathy's LLM Wiki pattern: knowledge is compiled once and kept current instead of being re-derived for every query. Kwipu takes the opposite route and queries a graph of notes at question time.
Version 0.6.13 shipped on October 22, 2026, the 62nd release since 0.1.0 on April 8. The GitHub release notes only point to the downloads; the changelog itself lives in the app's settings (Changelog) and in src/lib/changelog.ts.