MCP 服务器

RAG Context MCP Server

A lightweight server that provides persistent memory and context management for AI assistants using local vector storage and database, enabling efficient storage and retrieval of contextual information through semantic search and indexed retrieval.

README

RAG Context MCP Server

A lightweight Model Context Protocol (MCP) server that provides persistent memory and context management using local vector storage and database. This server enables AI assistants to store and retrieve contextual information efficiently using both semantic search and indexed retrieval.

Features

Local Vector Storage: Uses Vectra for efficient vector similarity search
Persistent Memory: SQLite database for reliable data persistence
Semantic Search: Automatic text embedding using Xenova/all-MiniLM-L6-v2 model
Hybrid Retrieval: Combines semantic search with indexed database queries
Simple API: Just two tools - setContext and getContext
Lightweight: Minimal dependencies, runs entirely locally
Privacy-First: All data stored locally, no external API calls

Installation

Using npm

npm install -g @rag-context/mcp-server

Using npx (no installation required)

npx @rag-context/mcp-server

Configuration

For Claude Desktop

Add the following to your Claude Desktop configuration file:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "rag-context": {
      "command": "npx",
      "args": ["@rag-context/mcp-server"],
      "env": {
        "RAG_CONTEXT_DATA_DIR": "/path/to/your/data/directory"
      }
    }
  }
}

For Cursor

In Cursor settings, add the MCP server:

env RAG_CONTEXT_DATA_DIR=/path/to/your/data/directory npx @rag-context/mcp-server

Environment Variables

RAG_CONTEXT_DATA_DIR: Directory where the database and vector index will be stored (default: ~/.rag-context-mcp)

Usage

The server exposes two main tools:

setContext

Store information in memory with automatic vectorization:

{
  "tool": "setContext",
  "arguments": {
    "key": "user_preferences",
    "content": "The user prefers dark mode and uses VS Code as their primary editor",
    "metadata": {
      "category": "preferences",
      "timestamp": "2024-01-15"
    }
  }
}

getContext

Retrieve relevant context using semantic search:

{
  "tool": "getContext",
  "arguments": {
    "query": "What are the user's editor preferences?",
    "limit": 5,
    "threshold": 0.7
  }
}

System Prompt for AI Assistants

To effectively use this MCP server, add the following to your AI assistant's system prompt:

## Memory and Context Management

You have access to a persistent memory system through the RAG Context MCP server. This allows you to store and retrieve information across conversations.

### When to Store Context

Store information when:

- Users share preferences, settings, or personal information
- Important project details or configurations are discussed
- Key decisions or agreements are made
- Useful code snippets or solutions are created
- Learning about user's workflow, tools, or environment

### How to Store Context

Use the `setContext` tool with:

- A descriptive, unique key (e.g., "project_setup_nextjs", "user_pref_editor")
- Clear, concise content that captures the essential information
- Relevant metadata (category, project, date, etc.)

Example:

```json
{
  "key": "project_api_structure",
  "content": "The project uses a REST API with /api/v1 prefix. Authentication is handled via JWT tokens in the Authorization header. Main endpoints: /users, /posts, /comments",
  "metadata": {
    "project": "blog-platform",
    "type": "architecture",
    "date": "2024-01-15"
  }
}
```

When to Retrieve Context

Retrieve context when:

Starting a new conversation about a previously discussed topic
Users reference past discussions or decisions
You need to recall specific technical details or preferences
Building upon previous work or solutions

How to Retrieve Context

Use the getContext tool with:

A natural language query describing what you're looking for
Appropriate limit (usually 3-5 results)
Threshold of 0.7 for balanced precision/recall

Example:

{
  "query": "API authentication setup for the blog project",
  "limit": 3,
  "threshold": 0.7
}

Best Practices

Be Selective: Store important, reusable information, not every detail
Use Clear Keys: Make keys descriptive and searchable
Add Metadata: Include project names, categories, and dates
Update Existing: Use the same key to update information rather than creating duplicates
Query Naturally: Write queries as you would ask a colleague

Remember: This memory persists across all conversations, making you more helpful over time by remembering important context and user preferences.


## Architecture

The server uses a hybrid approach for optimal performance:

1. **SQLite Database**: Stores the actual content with metadata, provides fast key-based lookups
2. **Vector Index**: Enables semantic search using embeddings
3. **Local Embeddings**: Uses Xenova/transformers for privacy-preserving, local text embedding

## Data Storage

All data is stored locally in the specified data directory:

<RAG_CONTEXT_DATA_DIR>/ ├── memories.db # SQLite database └── vectors.index # Vectra vector index


## Development

### Building from Source

```bash
# Clone the repository
git clone https://github.com/yourusername/rag-context-mcp.git
cd rag-context-mcp

# Install dependencies
npm install

# Build the project
npm run build

# Run in development mode
npm run dev

Running Tests

npm test

Privacy and Security

All data is stored locally on your machine
No external API calls for embeddings (uses local model)
No telemetry or data collection
You control where data is stored via RAG_CONTEXT_DATA_DIR

Troubleshooting

Common Issues

"VectorStore not initialized" error
- Ensure the data directory exists and has write permissions
- Check that the RAG_CONTEXT_DATA_DIR path is valid
Slow first startup
- The embedding model is downloaded on first use (~30MB)
- Subsequent starts will be much faster
High memory usage
- The embedding model requires ~200MB RAM
- Consider limiting the number of stored contexts