Semlink

中文

Semlink is an Obsidian plugin that vectorizes your Vault notes and adds a conversational AI assistant directly in the sidebar — plus an MCP server so Claude Desktop, Claude Code, Cursor and other AI tools can search and read your notes.

Ask questions in natural language: Semlink retrieves the most relevant notes, feeds them to your chat model, and answers with a visible thinking process, tool calls and cited sources — all grounded in your own notes.

Features

• Semantic Search: query your Vault in natural language; returns the most relevant chunks by vector similarity
• Conversational Q&A: an in-sidebar chat panel where the LLM answers from retrieved notes, with collapsible thinking process, tool-call traces and numbered reference sources
• Tool-using agent: the model can search, grep, read notes and inspect your current open note while answering
• Context management: prior turns are sent as a native message array (stable prompt prefix → high cache hit rate); the history is trimmed by a sliding window once it approaches the context limit
• Real-time Indexing: file changes are picked up automatically; indexing yields while you are actively working
• MCP Server: expose search/read tools over HTTP (JSON-RPC) for AI clients
• Feishu (Lark) bot: scan a QR code to bind, then chat with your Vault from Feishu via streaming cards

Chat Panel

• Click the Semlink icon in the left ribbon (or run Semlink: Open Search)
• Type a question — the model first thinks, then calls tools if it needs more, then answers with sources below
• Drag & drop: drop notes ANYWHERE in the panel — a guidance overlay appears while dragging — to attach them to the input ([[links]] supported inline)
• History: every conversation is saved; open it from the menu button, first question shown as the header subtitle
• Model switcher: pick any model from any configured provider; the context-usage ring and cache hit rate are shown next to it
• Home question cards: a curated pool of example prompts (recent notes / the "reading" topic / summarize the current note / knowledge review / duplicate notes / monthly review), three picked at random per visit

Tools

MCP additionally exposes index_status and reindex for index management.

How It Works

Obsidian Vault Notes
        │
        ▼
   Text Chunking
        │
        ▼
  Embedding API (BGE-M3)  ──→  Vector Embedding
        │
        ▼
  Local Storage (SQLite + Binary)
        │
        ▼
  Chat Pipeline (retrieval → LLM with tools → answer)
        │                        │
        ▼                        ▼
  Sidebar Chat Panel        MCP HTTP Server (:3001)
                                │
                                ▼
                      Claude / Cursor / Other AI Clients

Installation

Option 1: Build from Source

git clone https://gitee.com/ouzhongyuan/semlink.git
cd semlink
npm install
npm run build
# Copy the whole directory to MyVault/.obsidian/plugins/semlink/

Option 2: Direct Download

Download main.js, manifest.json, styles.css from the Release page and place them in YourVault/.obsidian/plugins/semlink/.

Enable

1. Obsidian → Settings → Community plugins
2. Find Semlink and enable it

Configuration

Settings → Semlink:

Connect AI Clients

Client configs are auto-generated at the bottom of the plugin settings page.

Claude Desktop / Cursor

{
  "mcpServers": {
    "semlink": {
      "type": "http",
      "url": "http://127.0.0.1:3001/mcp"
    }
  }
}

With access key:

{
  "mcpServers": {
    "semlink": {
      "type": "http",
      "url": "http://127.0.0.1:3001/mcp",
      "headers": { "Authorization": "Bearer your-key" }
    }
  }
}

Claude Code

claude mcp add --transport http semlink http://127.0.0.1:3001/mcp
# with key:
claude mcp add --transport http semlink http://127.0.0.1:3001/mcp --header "Authorization: Bearer your-key"

Feishu (Lark) Bot

1. Create a bot app in the Feishu open platform (enable bot capability, grant im:message / im:message:send_as_bot / cardkit:card:write etc., subscribe events via long connection)
2. In Settings → Semlink → Bot, enter the App ID / App Secret and scan the QR code with your Feishu
3. Send /bind <code> to the bot to finish binding
4. Chat with the bot — it replies with streaming cards showing the thinking process, tool calls and the answer

Commands

Data Storage

Everything stays local:

Never uploaded anywhere. Indexing consumes embedding API credits (e.g. SiliconFlow BGE-M3).

Embedding Models

Tech Stack

• Embedding: SiliconFlow API (BGE-M3, 1024-dim)
• Storage: sql.js (SQLite WASM, embedded in the bundle) + binary vector file
• Search: brute-force cosine similarity
• Chat: OpenAI / Anthropic-compatible chat APIs with streaming + tool calling
• MCP: HTTP transport (JSON-RPC 2.0)
• Bot: Feishu long-connection SDK + CardKit streaming cards

Development

npm run dev     # watch + auto-rebuild
npm run build   # production build

Notes

• Indexing consumes embedding API credits; watch usage on large Vaults
• Vector search runs in local memory; 100K+ notes may use significant memory
• MCP server binds to 127.0.0.1 by default (localhost only)

License

MIT