KnowledgeSmith MCP Server

MCP Server for Graphiti memory and document chunking. Previously included RBT document editing tools (now archived).

📦 Archive Notice

RBT Document Editor Tools (Archived 2025-10-09)

The RBT document editing功能已於 2025-10-09 封存，改用原生 Claude Code Read/Edit/Write 工具以降低維護成本和 token 使用。

封存內容：

• document_service.py - 文件服務

• document_parser.py - 文件解析器

• 11 個 editor MCP 工具（get_outline, read_content, update_block 等）

• templates/ - 文件模板

• cache.py - 文件快取

保留功能：

• ✅ chunking/ - 文件分塊與同步功能

• ✅ graphiti_tools.py - Graphiti 記憶體功能（8 個工具）

如何恢復封存的代碼：

# 查看封存版本
git show v-with-editor

# 恢復特定檔案
git checkout v-with-editor -- rbt_mcp_server/document_service.py

# 或建立分支使用完整封存版本
git checkout -b restore-editor v-with-editor

🎯 Current Features

Graphiti Knowledge Graph Integration

• Intelligent Chunking: Automatically split documents into semantic chunks based on document structure (sections for RBT, H3 headings for Markdown)

• Incremental Sync: Only update changed chunks, preserving unchanged content

• Neo4j Backend: Store document chunks as episodes in Graphiti knowledge graph

• graphiti-memory Compatible: Drop-in replacement with same search_nodes/search_facts API

• 8 MCP Tools: add_document, search_memory_nodes, search_memory_facts, get_episodes, delete_episode, get_entity_edge, delete_entity_edge, clear_graph

📦 Installation

Prerequisites

1. Setup Neo4j Database

Using Docker (recommended):

docker run \
  -p 7474:7474 \
  -p 7687:7687 \
  -e NEO4J_AUTH=neo4j/your-password \
  --name neo4j \
  neo4j:latest

Verify at: http://localhost:7474

2. Get OpenAI API Key

Required for Graphiti embeddings and graph operations.

Install MCP Server

Option 1: Install from source (uv)

# Clone repository
git clone https://github.com/yourusername/KnowledgeSmith.git
cd KnowledgeSmith

# Install with uv
uv pip install -e .

Option 2: Direct installation

uv pip install rbt-mcp-server

🚀 Quick Start

1. Configure Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "graphiti-memory-server": {
      "type": "stdio",
      "command": "rbt-mcp-server",
      "env": {
        "RBT_ROOT_DIR": "/path/to/your/document/root",
        "NEO4J_URI": "bolt://localhost:7687",
        "NEO4J_USER": "neo4j",
        "NEO4J_PASSWORD": "your-password",
        "OPENAI_API_KEY": "your-openai-api-key"
      }
    }
  }
}

Required Environment Variables:

• RBT_ROOT_DIR: Root directory for document comparison (required for add_document tool)

• NEO4J_URI, NEO4J_USER, NEO4J_PASSWORD: Neo4j database connection

• OPENAI_API_KEY: OpenAI API key for Graphiti embeddings

Or use full uv command:

{
  "mcpServers": {
    "graphiti-memory-server": {
      "type": "stdio",
      "command": "uv",
      "args": ["run", "rbt-mcp-server"],
      "env": {
        "RBT_ROOT_DIR": "/path/to/your/document/root",
        "NEO4J_URI": "bolt://localhost:7687",
        "NEO4J_USER": "neo4j",
        "NEO4J_PASSWORD": "your-password",
        "OPENAI_API_KEY": "your-openai-api-key"
      }
    }
  }
}

2. Set Environment Variables (Optional - if not using Claude Desktop)

# Required for add_document tool
export RBT_ROOT_DIR=/path/to/your/document/root

# Required for Graphiti integration
export NEO4J_URI=bolt://localhost:7687
export NEO4J_USER=neo4j
export NEO4J_PASSWORD=your-password
export OPENAI_API_KEY=your-openai-api-key

3. Test the Server

rbt-mcp-server

📚 Available MCP Tools

Document Management

1. add_document - Sync documents to knowledge graph with automatic chunking

   • Supports Markdown (chunked by H3 headings) and RBT documents

   • Incremental sync: only updates changed chunks

Knowledge Graph Query

2. search_memory_nodes - Search knowledge graph nodes (entities, preferences, procedures)

3. search_memory_facts - Search knowledge graph facts (relationships)

4. get_episodes - Retrieve recent memory episodes

Data Management

5. delete_episode - Delete specific episode

6. get_entity_edge - Get entity relationship edge by UUID

7. delete_entity_edge - Delete entity relationship edge

8. clear_graph - Clear all data from knowledge graph (⚠️ irreversible)

🔗 Graphiti Integration Usage

Adding Documents to Knowledge Graph

General Markdown Documents:

add_document(
    new_file_path="/absolute/path/to/document.md",
    project_id="my-project",
    file_path="docs/guide.md"  # relative path for general docs
)

RBT Documents (REQ/BP/TASK):

add_document(
    new_file_path="/absolute/path/to/TASK-001.md",
    project_id="knowledge-smith",
    feature_id="my-feature",
    rbt_type="TASK",
    file_path="001"  # task number for TASK documents
)

Searching Knowledge

# Search for nodes (entities, preferences, procedures)
results = await search_nodes(
    query="documentation preferences",
    group_ids=["knowledge-smith"],
    entity="Preference",
    max_nodes=10
)

# Search for facts (relationships)
facts = await search_facts(
    query="task dependencies",
    group_ids=["knowledge-smith"],
    max_facts=10
)

Difference from graphiti-memory MCP

This MCP server extends the original graphiti-memory MCP with document chunking capabilities:

• Original graphiti-memory: Stores entire documents as single episodes

• This MCP (graphiti-chunk-mcp): Automatically chunks documents into semantic sections

  • RBT documents: Split by section (sec-*)

  • Markdown documents: Split by H3 headings (###)

  • Incremental updates: Only sync changed chunks

API Compatibility: All search_nodes, search_facts, get_episodes functions maintain the same interface as graphiti-memory.

📖 Documentation

• Markdown Writing Guide - How to write Markdown documents that work well with the chunker

• MCP Installation Guide - Detailed installation and usage instructions

🧪 Development

Install development dependencies:

uv sync --dev

Run tests:

RBT_ROOT_DIR=/test/root uv run pytest -v

Test coverage:

RBT_ROOT_DIR=/test/root uv run pytest --cov=rbt_mcp_server --cov-report=html

📝 License

MIT License

🤝 Contributing

Contributions welcome! Please open an issue or submit a pull request.