LLM Wiki — Desktop App Turns Documents Into Self-Updating Knowledge Base

LLM Wiki is a cross-platform desktop application that automatically builds and maintains a persistent, interlinked knowledge base from user documents. Unlike traditional retrieval-augmented generation (RAG), which re-analyzes source material on every query, LLM Wiki compiles knowledge once into a structured wiki and keeps it current as new documents arrive.

The system's architecture consists of three layers, adapted from Karpathy's LLM Wiki pattern: Raw Sources (immutable original documents), Wiki (LLM-generated pages and indexes), and Schema (structural rules and configuration). Users can ingest multiple document formats—PDF, Office documents, EPUB/MOBI, Org mode, images, media, web clips, and batches of URLs—via a persistent queue that survives application restarts. A SHA256 incremental cache ensures that unchanged files are skipped, saving LLM tokens. Folder imports preserve directory structure and pass context to the LLM for better classification. The app automatically detects external changes in the raw/sources/ folder and syncs them with ingest or deletion, keeping the system in sync without manual intervention.

The ingest process uses a two-step chain-of-thought approach. In Step 1, the LLM analyzes the source material to identify key entities, concepts, arguments, and connections to existing wiki content, as well as contradictions or gaps. In Step 2, the LLM generates wiki files—source summaries with YAML frontmatter linking back to original sources, entity and concept pages with cross-references, and updated index and log files. The system also flags items for human review and can suggest web research queries to fill gaps. Every generated page includes a sources[] field in its frontmatter, ensuring full source traceability.

The application features a three-column desktop interface: a Knowledge Tree and File Tree on the left, a Chat panel in the center, and a Preview pane on the right. An icon sidebar switches between Wiki, Sources, Search, Graph, Lint, Review, Deep Research, and Settings views. All state—conversations, settings, review items, and project configuration—persists across restarts. Users can select from predefined scenario templates (Research, Reading, Personal Growth, Business, General) that pre-configure purpose.md (the wiki's goals and scope) and schema.md (structural rules).

LLM Wiki includes a knowledge graph with advanced relevance modeling. The system applies a 4-signal relevance model that scores connections between pages: direct [[wikilinks]] receive a ×3.0 weight, shared raw sources receive ×4.0, common neighbors (Adamic-Adar) receive ×1.5, and same-type pages receive ×1.0 bonus. Graph visualization uses sigma.js, graphology, and ForceAtlas2 layout; node colors represent page types or discovered communities, node sizes reflect link count (√ scaled), and edge thickness and color indicate relevance weight. Louvain community detection automatically discovers knowledge clusters based on link topology and scores each cluster by cohesion (intra-edge density); clusters with cohesion below 0.15 are flagged as sparse. The Graph Insights feature detects surprising connections (cross-community edges, peripheral-to-hub couplings) and knowledge gaps (isolated pages, sparse communities, bridge nodes connecting multiple clusters), ranking them by a composite surprise score. Users can click any insight to highlight the corresponding nodes and edges; knowledge gaps and bridge nodes have a one-click Deep Research button.

Query retrieval uses a multi-phase pipeline. Phase 1 applies tokenized search (word splitting and stop-word removal in English; CJK bigram tokenization in Chinese), with title matches receiving a +10 score bonus. Phase 1.5 optionally adds vector semantic search via any OpenAI-compatible /v1/embeddings endpoint, storing embeddings in LanceDB (Rust backend) for fast approximate nearest neighbor retrieval; semantic results are merged with keyword matches. Phase 2 applies graph expansion, using top search results as seed nodes and following 2-hop connections ranked by the 4-signal relevance model. Phase 3 controls context budget, with configurable token limits (4K to 1M) and proportional allocation across wiki pages (60%), chat history (20%), index (5%), and schema (15%).

The app includes a Rust backend chat agent that can read wiki, source, and graph data, perform web retrieval, generate workspace files, and execute shell commands with user approval. Local "skills" (SKILL.md files) can be scanned and enabled; users select skills with /skill and the agent reads instructions on demand. Generated outputs—Markdown, HTML, images, and other files—appear in an Outputs preview with quick folder access. Mermaid diagram code blocks render directly in chat and preview, with compact error cards instead of raw parser output. An async review system flags items for human judgment with predefined actions and pre-generated search queries.

A Chrome web clipper enables one-click web page capture with auto-ingest into the knowledge base. The system exposes a local HTTP API on 127.0.0.1:19828 for JSON queries, a bundled MCP (Model Context Protocol) server for hybrid search and file operations, and a ready-made agent skill that installs into Claude Code or Codex with one command (npx skills add …). All state is portable: users can export and import complete project archives across devices and rebuild the wiki index from existing pages. The project is cross-platform (desktop) and supports multiple LLM provider configurations, allowing independent routing of Chat and Ingest calls to different models with custom headers and streaming output.

LLM Wiki adapts Andrej Karpathy's LLM Wiki pattern—a methodology for using LLMs to build and maintain personal knowledge bases—into a fully-featured desktop application. The original pattern is abstract and relies on an LLM agent to implement; this project is a concrete, cross-platform implementation with substantial enhancements. The core architecture remains faithful: immutable raw sources feed into an LLM-generated wiki, which is governed by structural rules and purpose statements. However, the application layer transforms how that architecture is used.

The most significant innovation is the two-step chain-of-thought ingest process. Rather than analyzing and generating wiki pages in a single LLM pass, the system splits the work into an analysis phase (identifying entities, concepts, connections, and contradictions) followed by a generation phase (producing wiki files with source citations). This separation, combined with SHA256 incremental caching and a persistent queue, reduces token waste and makes the system resilient to interruptions—a key advantage for large or evolving document collections.

A second major addition is the knowledge graph with relevance scoring and community detection. The original pattern uses [[wikilinks]] for cross-references but does not visualize or analyze the resulting graph structure. LLM Wiki applies a weighted 4-signal model (direct links, source overlap, shared neighbors, type affinity) to rank page relationships, then uses Louvain clustering to discover natural knowledge areas. The system then proactively surfaces "surprising connections" (cross-cluster edges that may indicate novel insights) and "knowledge gaps" (isolated or sparsely-connected pages that may need research or deeper coverage). This transforms the wiki from a static repository into an actionable intelligence tool.

The chat agent and Deep Research feature extend the system beyond passive storage. Users can query the wiki with full graph context, request the LLM to fill knowledge gaps via web search (using Tavily, SerpApi, or SearXNG), and auto-ingest the results back into the wiki. The agent can read local "skill" files and execute approved workspace commands, making the system programmable within a larger workflow.

No comments yet. Be the first to share your thoughts!