Overview
nashsu/llm_wiki is the most popular LLM Wiki implementation on GitHub with 14.3k stars. It is a Tauri desktop application that runs on Windows, Mac, and Linux, featuring a 4-signal relevance model and automatic document clustering.
Key Facts
- Platform: Tauri desktop app (cross-platform: Windows, Mac, Linux)
- License: GNU LGPL v3
- Stars: 16k, 1.9k forks (most popular LLM Wiki repo as of Aug 2026)
- Note: Uses a proprietary format (not OKF standard), so pages are less portable to other tools. This is the key trade-off: highest popularity and most features, but closed to cross-tool exchange.
Features
Core Architecture
- Tauri v2 desktop app — cross-platform native GUI (macOS, Windows, Linux) with three-column layout
- Rust backend — chat agent runtime with tool-using capabilities, streaming, and shell approval
- Purpose.md — formal “why” layer for the wiki (goals, key questions, research scope) read by LLM during every operation
- Three-layer architecture — raw sources (immutable) → wiki (LLM-generated) → schema (rules & config), faithful to Karpathy’s pattern
Ingest & Processing
- Two-step Chain-of-Thought ingest — LLM analyzes first (entities, concepts, contradictions), then generates wiki pages with source traceability
- SHA256 incremental cache — source content hashed before ingest; unchanged files skipped automatically
- Persistent ingest queue — serial processing with crash recovery, cancel, retry, and progress visualization
- Multi-format document parsing — PDF, DOCX, PPTX, XLSX, EPUB/MOBI, images, web clips, batches of URLs
- Folder import — recursive import preserving directory structure, folder path as LLM classification context
- Source folder auto-watch — detects external changes in raw/sources/ and keeps ingest/delete cleanup in sync
- Multimodal image ingestion — extract images from PDFs, generate captions with vision LLM, surface in search with lightbox preview
Search & Retrieval
- 4-signal knowledge graph — relevance model with direct links (×3.0), source overlap (×4.0), Adamic-Adar (×1.5), type affinity (×1.0)
- Vector semantic search — optional embedding-based retrieval via LanceDB, supports any OpenAI-compatible endpoint
- Multi-phase retrieval pipeline — tokenized search → vector search → graph expansion → budget control (4K-1M tokens)
- Source-grounded retrieval — “Read Sources Only” mode answers exclusively from original imported material
Knowledge Graph
- Sigma.js + ForceAtlas2 visualization — interactive graph with node colors by type/community, edge thickness by relevance weight
- Louvain community detection — automatic knowledge cluster discovery with cohesion scoring and 12-color palette
- Graph insights — surprising connections (cross-community edges) and knowledge gaps (isolated pages, sparse communities)
- Deep Research from graph — one-click research from knowledge gaps with LLM-optimized domain-aware topics
MCP / Tools
- MCP server — built-in Model Context Protocol server for hybrid search, file read, graph traversal, source rescan
- Local HTTP API — JSON API at 127.0.0.1:19828 with token protection for external tool integration
- Agent Skills — scan and enable local SKILL.md folders, select skills with /skill, agent reads skill instructions on demand
- Chrome Web Clipper — Manifest V3 extension with Readability.js extraction, auto-ingest into knowledge base
Quality & Governance
- Async review system — LLM flags items for human judgment with predefined actions (Create Page, Deep Research, Skip)
- Lint reports — orphan pages, broken links, missing entity pages, data gaps with suggested sources
- Deletion cascade — 3-method matching finds related wiki pages; shared entity preservation across multiple sources
Integration
- Obsidian compatibility — wiki directory works as an Obsidian vault with auto-generated .obsidian/ config
- Multi-conversation chat — independent sessions with persistence, cited references panel, regenerate, save-to-wiki
- Deep Research — web search via Tavily, SerpApi, or SearXNG with multi-query synthesis and auto-ingest
- Mermaid + KaTeX rendering — inline diagrams and math formulas in chat and preview
- Generated Outputs Preview — visual preview of generated content (diagrams, tables, code)
- Scenario templates — pre-built templates for common wiki use cases
- i18n — English + Chinese interface (react-i18next)
Data & Storage
- LanceDB embedded vector DB — optional, Rust-backed, fast ANN retrieval
- Project management — export/import complete project archives, rebuild wiki index from existing pages
- Configurable context window — slider from 4K to 1M tokens with proportional budget allocation
- Flexible model configuration — supports any OpenAI-compatible endpoint for LLM and embeddings
Relationship to This Vault
nashsu is the reference for: the “most popular LLM Wiki with a GUI” position in knowledge-management-approaches-vs-llm-wiki and llm-wiki-vs-memory-and-graph-rag. The vault chose an OKF-compliant markdown wiki instead for portability, but nashsu’s MCP server and Rust backend demonstrate production-grade features this vault targets for Fase C.
Comparison with Other Implementations
Among the major LLM Wiki tools, nashsu is the most popular but the least standards-compliant:
| Feature | nashsu | atomicstrata | lucasastorian |
|---|---|---|---|
| OKF Compliant | No | Yes | No |
| Desktop GUI | Yes | No | No |
| MCP Server | Yes | Yes | Yes |
| License | LGPL | MIT | Not specified |
This makes it the best choice for a standalone desktop wiki experience, while atomicstrata-llm-wiki-compiler is better when interoperability and portability across tools are important.
Relationships
- llm-wiki — implemented-by: Most popular implementation of the LLM Wiki pattern
- open-knowledge-format — contradicts: Does not use OKF (uses proprietary format instead)
- lucasastorian-llmwiki — sibling-of: Alternative full-featured implementation (full-stack instead of desktop)
- atomicstrata-llm-wiki-compiler — sibling-of: Standards-compliant alternative
🔗 Conexão inferida — sugestão do segundo cérebro
Liga para: llm-wiki Evidência observável: Nashsu built an LLM Wiki using the pattern described by Karpathy; documented in the community discussion.