llm_wiki

Product Introduction

1. Definition: LLM Wiki is a cross-platform desktop application that functions as an AI-powered, persistent knowledge base system. It is a concrete implementation of the "LLM Wiki" architectural pattern, designed to transform unstructured documents into a structured, interlinked wiki using Large Language Models (LLMs).
2. Core Value Proposition: It exists to solve the inefficiency of traditional Retrieval-Augmented Generation (RAG) systems by moving from a stateless, query-time retrieval model to a stateful, compiled knowledge model. Instead of re-analyzing documents for every question, LLM Wiki incrementally builds and maintains a permanent, organized wiki, enabling faster, more consistent, and source-grounded knowledge access for researchers, writers, and knowledge workers.

Main Features

1. Two-Step Chain-of-Thought Ingest: This feature decouples analysis from generation for higher quality. First, an LLM performs a structured analysis of a source document, identifying key entities, concepts, and connections. Second, a separate LLM call uses this analysis to generate or update wiki pages. It incorporates SHA256-based incremental caching to skip unchanged files, a persistent ingest queue with crash recovery, and automatic embedding for vector search.
2. 4-Signal Knowledge Graph with Louvain Community Detection: Beyond simple wikilinks, LLM Wiki constructs a quantitative knowledge graph. A 4-signal relevance model (Direct Links, Source Overlap, Adamic-Adar, Type Affinity) weights connections between pages. The Louvain algorithm automatically detects knowledge clusters (communities) within this graph, visualized with a sigma.js-powered interface. The system analyzes this graph to surface "Graph Insights" like surprising cross-community connections and knowledge gaps (isolated pages, sparse clusters).
3. Rust Backend Chat Agent with Skills & MCP Server: Queries are processed by a tool-using AI agent runtime built in Rust, not just a simple LLM call. This agent can search the wiki, sources, graph, and web (via Tavily/SerpApi), read files, and execute approved shell commands. It supports local "SKILL.md" files that the agent can read on demand. The system includes a built-in MCP (Model Context Protocol) server and local HTTP API (127.0.0.1:19828), allowing external AI agents like Claude Code to directly query the knowledge base.
4. Optimized Hybrid Retrieval Pipeline: The query system employs a multi-phase retrieval strategy. It begins with tokenized full-text search (with CJK bigram support for Chinese). Optionally, it integrates vector semantic search via LanceDB and any OpenAI-compatible embeddings endpoint. Results are then expanded via the 4-signal knowledge graph. Finally, a context budget manager allocates tokens proportionally (60% wiki, 20% chat history, etc.) before assembling the final prompt sent to the LLM.
5. Deep Research & Asynchronous Review System: For identified knowledge gaps, the "Deep Research" feature lets an LLM generate optimized search topics and queries, performs multi-query web searches, and auto-ingests results. The "Review System" creates an async human-in-the-loop workflow; during ingest, the LLM flags items needing human judgment (e.g., "Create Page," "Research") with pre-generated actions and search queries, which users can address later in a dedicated panel.

Problems Solved

1. Pain Point: The inefficiency and inconsistency of stateless RAG. Traditional RAG re-processes documents from scratch for each query, wasting computational resources and LLM context, and can produce varying answers.
2. Target Audience: Academic researchers, technical writers, investigative journalists, product managers, and legal professionals who need to synthesize information from large, evolving document collections (PDFs, research papers, internal reports, web clips).
3. Use Cases: Literature review for a PhD thesis, maintaining a competitive intelligence wiki, building a personal knowledge management (PKM) system from diverse reading materials, onboarding into a complex technical project by digesting its documentation.

Unique Advantages

1. Differentiation: Unlike note-taking apps (Obsidian, Logseq) which are passive containers, LLM Wiki actively structures content. Unlike cloud-based AI notebooks, it is a private desktop application. Unlike classic RAG pipelines (e.g., using LangChain), it builds a persistent, queryable knowledge artifact rather than performing one-time retrieval.
2. Key Innovation: The synthesis of the "LLM Wiki" persistent knowledge base pattern with a production-grade desktop application stack (Tauri, Rust, React, TypeScript), enhanced with a quantitative knowledge graph, a tool-using agent runtime, and a flexible hybrid retrieval system. The "Purpose.md" file guiding LLM context and the "Schema.md" enforcing structure create a coherent, goal-directed knowledge system.

Frequently Asked Questions (FAQ)

1. How does LLM Wiki handle data privacy and where is my data stored? All processing occurs locally on your machine. Documents, the generated wiki, and the vector database (if used) are stored in your local project directory. API keys for LLMs and search services are stored locally and only transmit data to the respective providers you configure.
2. What document formats can LLM Wiki parse and ingest? It supports PDF, Microsoft Office documents (DOCX, PPTX), EPUB, MOBI, Org mode files, images (with OCR/Vision LLM captioning), media files, and web clips. It includes built-in and configurable (local/cloud) parsers like MinerU for complex PDFs.
3. Can I use LLM Wiki with open-source or local LLMs instead of OpenAI? Yes. It features flexible model configuration per project, allowing you to route "Chat" and "Ingest" tasks to different endpoints. It supports any OpenAI-compatible API (e.g., LM Studio, Ollama, vLLM) and can manage custom providers, headers, and streaming.
4. Is the generated wiki compatible with other tools like Obsidian? Yes. The wiki output uses standard Markdown with [[wikilink]] syntax and YAML frontmatter, making it directly usable as an Obsidian vault. The application is designed for interoperability.
5. How does the "Deep Research" feature work to fill knowledge gaps? When the graph analysis identifies a gap (e.g., an isolated page), you can trigger Deep Research. The LLM reads your purpose.md and overview.md to generate a domain-aware research topic and search engine queries. It then uses configured providers (Tavily, SerpApi, SearXNG) to fetch full web content, which an LLM synthesizes into a research page and auto-ingests into the wiki.