WHAT IS IT?
LLM Wiki is a desktop application that reads your documents and builds a structured, cross-linked wiki out of them, one page at a time. Where classic RAG answers from scratch on every query, here the LLM compiles knowledge once and updates it as new sources arrive. What you get is a Markdown folder that opens as-is in Obsidian, not an opaque vector store.
WHY IS IT INTERESTING?
- Two-step ingest: the LLM first analyzes the source (entities, concepts, contradictions with existing content), then writes the wiki pages. Every page carries a
sources[]field in its YAML frontmatter that points back to the original document. - Weighted knowledge graph: four relevance signals, direct links (x3), shared sources (x4), Adamic-Adar (x1.5) and type affinity (x1). Louvain community detection then finds thematic clusters and scores their cohesion.
- Graph insights: the app surfaces cross-community connections, isolated pages, weakly cohesive clusters and bridge nodes. Any detected gap kicks off a one-click web search through Tavily, SerpApi or SearXNG.
- Wide format coverage: PDF via a built-in Rust parser or MinerU, DOCX, PPTX, XLSX, EPUB/MOBI, Org mode, images, web clips. A SHA256 cache skips unchanged files so you don't pay for the same tokens twice.
- Local HTTP API and MCP server: everything is exposed on
127.0.0.1:19828behind a token. A ready-made skill wires Claude Code or Codex into the wiki with a singlenpx skills addcommand. - Optional vector search: LanceDB in the Rust backend, with any OpenAI-compatible endpoint for embeddings. It is off by default, and the pipeline then falls back to tokenized search plus graph expansion.
USE CASES
- Digest a corpus of academic PDFs into a browsable knowledge base with traceability back to sources
- Archive your web reading through the Chrome clipper, where each captured page triggers automatic ingest
- Wire a Claude Code agent into your own knowledge base through the MCP server
- Cross-read several internal reports and spot the blind spots the graph flags as knowledge gaps