LLM Wiki is a desktop application for macOS, Windows, and Linux that has a language model read your documents and build a structured, interlinked wiki from them. Instead of retrieving and answering from raw sources on every query, the model compiles knowledge once and keeps the wiki current as sources change. It implements Andrej Karpathy's LLM Wiki pattern: immutable raw sources, a model-generated wiki, and a schema of rules, with ingest, query, and lint as the three core operations.
Ingest runs in two model calls: the first analyses a source for entities, concepts, contradictions, and links to existing pages, and the second writes the wiki pages, each recording which sources contributed to it. Unchanged files are skipped by content hash, and a persistent queue processes sources one at a time with retry and crash recovery. A purpose document states the wiki's goals and research scope, and the model reads it on every ingest and query.
Queries combine tokenized keyword search, optional vector search, and graph expansion over the wiki's links, within a configurable context budget. The wiki directory also works as an Obsidian vault. A Rust chat agent answers over the wiki, sources, graph, and web, and a local HTTP API plus a bundled MCP server let external agents such as Claude Code and Codex search and read the wiki.
The app is built with Tauri and React, and it works with OpenAI, Anthropic, Google, Ollama, and custom providers, with chat and ingest routed to separate models if needed.
Features
- Two-step ingest: an analysis pass followed by a generation pass, with source traceability on every page and an incremental cache
- Document parsing: PDF, Office documents, EPUB and MOBI, Org mode, images, media, web clips, and batches of URLs, with optional MinerU processing for complex PDFs
- Multimodal images: extracts images from PDFs, captions them with a vision model, and surfaces them in search results
- Knowledge graph: a relevance model over direct links, source overlap, Adamic-Adar, and type affinity, drawn as an interactive graph
- Community detection: Louvain clustering with cohesion scores to find knowledge clusters and weakly linked areas
- Graph insights: flags surprising connections, isolated pages, sparse communities, and bridge nodes, each with a Deep Research action
- Deep Research: generates search topics, queries Tavily, SerpApi, or SearXNG, and ingests the results into the wiki
- Vector search: optional embedding retrieval stored in LanceDB, with any OpenAI-compatible embeddings endpoint
- Source-grounded answers: a Read Sources Only mode answers solely from the original imported material
- Review queue: the model flags items for human judgment with predefined actions and pre-generated search queries
- Agent access: a token-protected local HTTP API, an MCP server, and an agent skill for Claude Code and Codex
- Web clipper: a Chrome extension captures pages and ingests them into the wiki
