Synapse
The self-hosted LLM wiki that turns your sources into connected knowledge.
Connect everything.
Drop in raw documents — Synapse reads them, writes interlinked wiki pages,
builds a weighted knowledge graph, and lets you chat with your own knowledge base, with citations.
Based on the LLM Wiki pattern by Andrej Karpathy (nashsu/llm_wiki), re-engineered for performance and self-hosting.
What it does
| |
|---|
| 📥 Ingest anything | Markdown, PDF, DOCX, PPTX, XLSX — via file watcher, upload, scheduled folder import, or the Chrome web clipper. An LLM analyzes each source and writes typed, interlinked wiki pages. |
| 🕸️ Knowledge graph | 4-signal edge weighting, server-side ForceAtlas2 layout, Louvain communities, WebGL rendering (sigma.js). Never freezes your browser. |
| 💬 Chat with citations | 4-phase retrieval (vector → graph expansion → token budget → assembly) over your wiki, streamed answers with [n] citations, save-to-wiki. |
| 🔍 Deep research | Bounded agentic loop over SearXNG web search → synthesized, cited pages, auto-ingested. |
| 🧑⚖️ Human in the loop | Review queue for AI proposals, bounded lint-fix loop with human-gated apply, cascade delete with dry-run. |
| 🔌 Bring your own AI | Pluggable inference provider: Local (Ollama), API (Anthropic / OpenAI-compatible), CLI (claude-agent-sdk). Switch per vault or per operation. |
| 📦 Obsidian-compatible | The wiki/ folder is a valid Obsidian vault: YAML frontmatter, [[wikilinks]], auto-generated .obsidian/. |
| 🤖 MCP server | Nine read/write tools (search, get/write page, graph neighborhood, review queue, ...) exposed over stdio and HTTP — talk to your wiki from Claude Desktop, Claude Code, or any MCP client. |
| 🖥️ Desktop app | Native macOS/Windows app (Tauri v2, ~6 MB) with auto-update from GitHub Releases. Point it at your server and go. |
| 📱 iOS app | Native SwiftUI client (iOS 17+) — Home, Wiki, Chat, Graph, and review queue against your own server, brand-matched to the desktop UI. |
Screenshots
| Knowledge graph (dark) | Wiki + graph workspace |
|---|
| |
| Search with type filters | Chat |
|---|
| |
Quick start (server)
git clone https://github.com/Emanuele-Chiummo/llm-wiki-synapse.git
cd llm-wiki-synapse
cp .env.example .env # configure DB, Ollama/Qdrant endpoints, provider
docker compose up -d
Open http://localhost:8000 (API) — serve the frontend with make dev or any static host.
Requires: Docker, plus the services you already run — Ollama (inference/embeddings, bge-m3), Qdrant (vectors), optionally SearXNG (web search).
Desktop app
Download the latest .dmg (macOS) or .exe (Windows) from Releases.
First launch: enter your backend URL (local servers are auto-detected). Builds are unsigned for now — macOS: right-click → Open; Windows: "More info" → "Run anyway". The app checks GitHub for updates at startup and offers to install them.
iOS app
Native SwiftUI client, source in ios/. Not on the App Store yet: build and run it
yourself with Xcode (free Apple ID, 7-day install), or produce an unsigned .ipa with
ios/build-unsigned-ipa.sh to sideload via
AltStore or Sideloadly. TestFlight distribution
is wired up (ios/scripts/testflight.sh) but not yet live — it needs a paid Apple Developer
Program enrollment. Full instructions: ios/README.md.
Documentation
Stack
FastAPI · SQLAlchemy 2 · PostgreSQL 16 · Qdrant · bge-m3 · FastMCP — React 19 · Vite · CodeMirror 6 · sigma.js · Zustand — Tauri v2 · Swift/SwiftUI.
License
MIT © 2026 Emanuele Chiummo
