llm_wiki is a cross-platform desktop application built with Tauri v2 and React 19 that transforms static documents into a persistent, self-maintaining knowledge base. Moving beyond standard RAG, the system implements a two-step chain-of-thought ingestion process: an initial analysis phase identifies entities, concepts, and connections, followed by a generation phase that produces structured Markdown wiki pages with YAML frontmatter for source traceability. The architecture utilizes a three-layer model—Raw Sources, Wiki, and Schema—and supports diverse formats including PDF, Office, EPUB, and web clips via a Chrome extension. Technical sophistication is provided through a 4-signal relevance engine (Direct links, Source overlap, Adamic-Adar, and Type affinity) and Louvain community detection to identify knowledge clusters. Retrieval is handled via a multi-phase pipeline combining tokenized search, graph expansion, and optional vector semantic search using LanceDB, which has been shown to improve recall from 58.2% to 71.4%. Additionally, the system features a Rust-based Agent runtime capable of tool-use, deep research via Tavily or SerpApi, and provides a local HTTP API and MCP server to enable integration with external AI agents like Claude Code.