LLM Wiki is a cross-platform desktop application that turns your documents into an organized, interlinked knowledge base — automatically. Instead of traditional RAG (retrieve-and-answer from scratch every time), the LLM incrementally builds and maintains a persistent wiki from your sources. Knowledge is compiled once and kept current, not re-derived on every query.
This project is based on Karpathy's LLM Wiki pattern — a methodology for building personal knowledge bases using LLMs. We implemented the core ideas as a full desktop application with significant enhancements.
Features
- Two-Step Chain-of-Thought Ingest — LLM analyzes first, then generates wiki pages with source traceability and incremental cache
- 4-Signal Knowledge Graph — relevance model with direct links, source overlap, Adamic-Adar, and type affinity
- Louvain Community Detection — automatic knowledge cluster discovery with cohesion scoring
- Graph Insights — surprising connections and knowledge gaps with one-click Deep Research
- Vector Semantic Search — optional embedding-based retrieval via LanceDB, supports any OpenAI-compatible endpoint
- Persistent Ingest Queue — serial processing with crash recovery, cancel, retry, and progress visualization
- Folder Import — recursive folder import preserving directory structure, folder context as LLM classification hint
- Deep Research — LLM-optimized search topics, multi-query web search, auto-ingest results into wiki
- Async Review System — LLM flags items for human judgment, predefined actions, pre-generated search queries
- Chrome Web Clipper — one-click web page capture with auto-ingest into knowledge base
README
LLM Wiki
A personal knowledge base that builds itself.
LLM reads your documents, builds a structured wiki, and keeps it current.
Two-Step Chain-of-Thought Ingest — LLM analyzes first, then generates wiki pages with source traceability and incremental cache
Multimodal Image Ingestion — extract embedded images from PDFs, generate factual captions with a vision LLM, surface them in image-aware search results with lightbox preview and jump-to-source
4-Signal Knowledge Graph — relevance model with direct links, source overlap, Adamic-Adar, and type affinity
Louvain Community Detection — automatic knowledge cluster discovery with cohesion scoring
Graph Insights — surprising connections and knowledge gaps with one-click Deep Research
Vector Semantic Search — optional embedding-based retrieval via LanceDB, supports any OpenAI-compatible endpoint
Persistent Ingest Queue — serial processing with crash recovery, cancel, retry, and progress visualization
Folder Import — recursive folder import preserving directory structure, folder context as LLM classification hint
Source Folder Auto-Watch — detects external changes in raw/sources/ and keeps ingest/delete cleanup in sync
Deep Research — LLM-optimized search topics, multi-query web search via Tavily, SerpApi, or SearXNG, auto-ingest results into wiki
Async Review System — LLM flags items for human judgment, predefined actions, pre-generated search queries
Chrome Web Clipper — one-click web page capture with auto-ingest into knowledge base
What is this?
LLM Wiki is a cross-platform desktop application that turns your documents into an organized, interlinked knowledge base — automatically. Instead of traditional RAG (retrieve-and-answer from scratch every time), the LLM incrementally builds and maintains a persistent wiki from your sources. Knowledge is compiled once and kept current, not re-derived on every query.
This project is based on Karpathy's LLM Wiki pattern — a methodology for building personal knowledge bases using LLMs. We implemented the core ideas as a full desktop application with significant enhancements.
Credits
The foundational methodology comes from Andrej Karpathy's llm-wiki.md, which describes the pattern of using LLMs to incrementally build and maintain a personal wiki. The original document is an abstract design pattern; this project is a concrete implementation with substantial extensions.
What We Kept from the Original
The core architecture follows Karpathy's design faithfully:
Three-layer architecture: Raw Sources (immutable) → Wiki (LLM-generated) → Schema (rules & config)
Three core operations: Ingest, Query, Lint
index.md as the content catalog and LLM navigation entry point
log.md as the chronological operation record with parseable format
[[wikilink]] syntax for cross-references
YAML frontmatter on every wiki page
Obsidian compatibility — the wiki directory works as an Obsidian vault
Human curates, LLM maintains — the fundamental role division
What We Changed & Added
1. From CLI to Desktop Application
The original is an abstract pattern document designed to be copy-pasted to an LLM agent. We built it into a full cross-platform desktop application with:
Three-column layout: Knowledge Tree / File Tree (left) + Chat (center) + Preview (right)
Icon sidebar for switching between Wiki, Sources, Search, Graph, Lint, Review, Deep Research, Settings
Custom resizable panels — drag-to-resize left and right panels with min/max constraints
Activity panel — real-time processing status showing file-by-file ingest progress
All state persisted — conversations, settings, review items, project config survive restarts
Scenario templates — Research, Reading, Personal Growth, Business, General — each pre-configures purpose.md and schema.md
2. Purpose.md — The Wiki's Soul
The original has Schema (how the wiki works) but no formal place for why the wiki exists. We added purpose.md:
Defines goals, key questions, research scope, evolving thesis
LLM reads it during every ingest and query for context
LLM can suggest updates based on usage patterns
Different from schema — schema is structural rules, purpose is directional intent
3. Two-Step Chain-of-Thought Ingest
The original describes a single-step ingest where the LLM reads and writes simultaneously. We split it into two sequential LLM calls for significantly better quality:
Step 1 (Analysis): LLM reads source → structured analysis
- Key entities, concepts, arguments
- Connections to existing wiki content
- Contradictions & tensions with existing knowledge
- Recommendations for wiki structure
Step 2 (Generation): LLM takes analysis → generates wiki files
- Source summary with frontmatter (type, title, sources[])
- Entity pages, concept pages with cross-references
- Updated index.md, log.md, overview.md
- Review items for human judgment
- Search queries for Deep Research
Additional ingest enhancements beyond the original:
SHA256 incremental cache — source file content is hashed before ingest; unchanged files are skipped automatically, saving LLM tokens and time
Persistent ingest queue — serial processing prevents concurrent LLM calls; queue persisted to disk, survives app restart; failed tasks auto-retry up to 3 times
Source folder auto-watch — files added, edited, or deleted in raw/sources/ outside the app are picked up automatically and reuse the same ingest/delete lifecycle as in-app actions
Queue visualization — Activity Panel shows progress bar, pending/processing/failed tasks with cancel and retry buttons
Auto-embedding — when vector search is enabled, new pages are automatically embedded after ingest
Source traceability — every generated wiki page includes a sources: [] field in YAML frontmatter, linking back to the raw source files that contributed to it
overview.md auto-update — global summary page regenerated on every ingest to reflect the latest state of the wiki
Guaranteed source summary — fallback ensures a source summary page is always created, even if the LLM omits it
Language-aware generation — LLM responds in the user's configured language (English or Chinese)
Progressive Sources view — large source folders render progressively while scrolling, keeping big source collections responsive
4. Knowledge Graph with Relevance Model
The original mentions [[wikilinks]] for cross-references but has no graph analysis. We built a full knowledge graph visualization and relevance engine:
4-Signal Relevance Model:
Signal
Weight
Description
Direct link
×3.0
Pages linked via [[wikilinks]]
Source overlap
×4.0
Pages sharing the same raw source (via frontmatter sources[])
Adamic-Adar
×1.5
Pages sharing common neighbors (weighted by neighbor degree)
Type affinity
×1.0
Bonus for same page type (entity↔entity, concept↔concept)
Composite surprise score ranks the most noteworthy connections
Dismissable — mark connections as reviewed so they don't reappear
Knowledge Gaps:
Isolated pages (degree ≤ 1) — pages with few or no connections to the rest of the wiki
Sparse communities (cohesion < 0.15, ≥ 3 pages) — knowledge areas with weak internal cross-references
Bridge nodes (connecting 3+ clusters) — critical junction pages that hold multiple knowledge areas together
Interactive:
Click any insight card to highlight corresponding nodes and edges in the graph; click again to deselect
Knowledge gaps and bridge nodes have a Deep Research button — triggers LLM-optimized research with domain-aware topics (reads overview.md + purpose.md for context)
Research topic shown in editable confirmation dialog before starting — user can refine topic and search queries
7. Optimized Query Retrieval Pipeline
The original describes a simple query where the LLM reads relevant pages. We built a multi-phase retrieval pipeline with optional vector search and budget control:
Phase 1: Tokenized Search
- English: word splitting + stop word removal
- Chinese: CJK bigram tokenization (每个 → [每个, 个…])
- Title match bonus (+10 score)
- Searches both wiki/ and raw/sources/
Phase 1.5: Vector Semantic Search (optional)
- Embedding via any OpenAI-compatible /v1/embeddings endpoint
- Stored in LanceDB (Rust backend) for fast ANN retrieval
- Cosine similarity finds semantically related pages even without keyword overlap
- Results merged into search: boosts existing matches + adds new discoveries
Phase 2: Graph Expansion
- Top search results used as seed nodes
- 4-signal relevance model finds related pages
- 2-hop traversal with decay for deeper connections
Phase 3: Budget Control
- Configurable context window: 4K → 1M tokens
- Proportional allocation: 60% wiki pages, 20% chat history, 5% index, 15% system
- Pages prioritized by combined search + graph relevance score
Phase 4: Context Assembly
- Numbered pages with full content (not just summaries)
- System prompt includes: purpose.md, language rules, citation format, index.md
- LLM instructed to cite pages by number: [1], [2], etc.
Vector Search is fully optional — disabled by default, enabled in Settings with independent endpoint, API key, and model configuration. When disabled, the pipeline falls back to tokenized search + graph expansion. Benchmark: overall recall improved from 58.2% to 71.4% with vector search enabled.
8. Multi-Conversation Chat with Persistence
The original has a single query interface. We built full multi-conversation support:
Conversation sidebar — quick switching between topics
Per-conversation persistence — each conversation saved to .llm-wiki/chats/{id}.json
Configurable history depth — limit how many messages are sent as context (default: 10)
Cited references panel — collapsible section on each response showing which wiki pages were used, grouped by type with icons
Reference persistence — cited pages stored directly in message data, stable across restarts
Regenerate — re-generate the last response with one click (removes last assistant + user message pair, re-sends)
Save to Wiki — archive valuable answers to wiki/queries/, then auto-ingest to extract entities/concepts into the knowledge network
9. Thinking / Reasoning Display
Not in the original. For LLMs that emit `` blocks (DeepSeek, QwQ, etc.):
Streaming thinking — rolling 5-line display with opacity fade during generation
Collapsed by default — thinking blocks hidden after completion, click to expand
Visual separation — thinking content shown in distinct style, separate from the main response
10. KaTeX Math Rendering
Not in the original. Full LaTeX math support across all views:
KaTeX rendering — inline $...$ and block $$...$$ formulas rendered via remark-math + rehype-katex
Milkdown math plugin — preview editor renders math natively via @milkdown/plugin-math
Auto-detection — bare \begin{aligned} and other LaTeX environments automatically wrapped with $$ delimiters
Unicode fallback — 100+ symbol mappings (α, ∑, →, ≤, etc.) for simple inline notation outside math blocks
11. Review System (Async Human-in-the-Loop)
The original suggests staying involved during ingest. We added an asynchronous review queue:
LLM flags items needing human judgment during ingest
Predefined action types: Create Page, Deep Research, Skip — constrained to prevent LLM hallucination of arbitrary actions
Search queries generated at ingest time — LLM pre-generates optimized web search queries for each review item
User handles reviews at their convenience — doesn't block ingest
12. Deep Research
Not in the original. When the LLM identifies knowledge gaps:
Web search via Tavily, SerpApi, or SearXNG finds relevant sources with full content extraction (no truncation)
Provider-specific configuration — Tavily and SerpApi use independent API keys; SerpApi supports selectable engines, while SearXNG uses a configured instance URL and search categories
Multiple search queries per topic — LLM-generated at ingest time, optimized for search engines
LLM-optimized research topics — when triggered from Graph Insights, LLM reads overview.md + purpose.md to generate domain-specific topics and queries (not generic keywords)
User confirmation dialog — editable topic and search queries shown for review before research starts
LLM synthesizes findings into a wiki research page with cross-references to existing wiki
Thinking display — `` blocks shown as collapsible sections during synthesis, auto-scroll to latest content
Auto-ingest — research results automatically processed to extract entities/concepts into the wiki
Task queue with 3 concurrent tasks
Research Panel — dedicated sidebar panel with dynamic height, real-time streaming progress
13. Browser Extension (Web Clipper)
The original mentions Obsidian Web Clipper. We built a dedicated Chrome Extension (Manifest V3):
Mozilla Readability.js for accurate article extraction (strips ads, nav, sidebars)
Turndown.js for HTML → Markdown conversion with table support
Project picker — choose which wiki to clip into (supports multi-project)
Local HTTP API (port 19827, tiny_http) — Extension ↔ App communication
Auto-ingest — clipped content automatically triggers the two-step ingest pipeline
Clip watcher — polls every 3 seconds for new clips, processes automatically
Offline preview — shows extracted content even when app is not running
14. Multi-format Document Support
The original focuses on text/markdown. We support structured extraction preserving document semantics:
The original has no deletion mechanism. We added intelligent cascade deletion:
Deleting a source file removes its wiki summary page
3-method matching finds related wiki pages: frontmatter sources[] field, source summary page name, frontmatter section references
Shared entity preservation — entity/concept pages linked to multiple sources only have the deleted source removed from their sources[] array, not deleted entirely
Index cleanup — removed pages are purged from index.md
Wikilink cleanup — dead [[wikilinks]] to deleted pages are removed from remaining wiki pages
16. Configurable Context Window
Not in the original. Users can configure how much context the LLM receives:
Slider from 4K to 1M tokens — adapts to different LLM capabilities
Proportional budget allocation — larger windows get proportionally more wiki content
60/20/5/15 split — wiki pages / chat history / index / system prompt
17. Cross-Platform Compatibility
The original is platform-agnostic (abstract pattern). We handle concrete cross-platform concerns:
Path normalization — unified normalizePath() used across 22+ files, backslash → forward slash
Unicode-safe string handling — char-based slicing instead of byte-based (prevents crashes on CJK filenames)
macOS close-to-hide — close button hides window (app stays running in background), click dock icon to restore, Cmd+Q to quit
Windows/Linux close confirmation — confirmation dialog before quitting to prevent accidental data loss
Tauri v2 — native desktop on macOS, Windows, Linux
GitHub Actions CI/CD — automated builds for macOS (ARM + Intel), Windows (.msi), Linux (.deb / .AppImage)
18. Other Additions
i18n — English + Chinese interface (react-i18next)
Settings persistence — LLM provider, API key, model, context size, language saved via Tauri Store
Obsidian config — auto-generated .obsidian/ directory with recommended settings
Markdown rendering — GFM tables with borders, proper code blocks, wikilink processing in chat and preview
Multi-provider LLM support — OpenAI, Anthropic, Google, Ollama, Custom — each with provider-specific streaming and headers
15-minute timeout — long ingest operations won't fail prematurely
dataVersion signaling — graph and UI automatically refresh when wiki content changes
# Prerequisites: Node.js 20+, Rust 1.70+
git clone https://github.com/nashsu/llm_wiki.git
cd llm_wiki
npm install
npm run tauri dev # Development
npm run tauri build # Production build
Chrome Extension
Open chrome://extensions
Enable "Developer mode"
Click "Load unpacked"
Select the extension/ directory
Quick Start
Launch the app → Create a new project (choose a template)
Go to Settings → Configure your LLM provider (API key + model)
Optional: configure Web Search providers and source folder auto-watch in Settings
Go to Sources → Import documents (PDF, DOCX, MD, etc.)
Watch the Activity Panel — LLM automatically builds wiki pages
