MCP 服务器

kavi-research-assistant-mcp

Enables AI to save, organize, search, and synthesize research materials using a local vector database with support for both OpenAI and Ollama backends.

README

KAVI RESEARCH

Your Premium AI Research Librarian

Features • Installation • Configuration • Usage • Contributing

</div>

🚀 Overview

KAVI RESEARCH is a premium Model Context Protocol (MCP) server designed to transform your AI into a dedicated research assistant.

Stop losing track of important findings. KAVI RESEARCH enables your AI to save, organize, search, and synthesize high-volume research materials using a local vector database. Whether you are using OpenAI or local Ollama models, KAVI RESEARCH keeps your knowledge accessible, private, and secure.

Newly Added in v2.1: Large Document Support! KAVI RESEARCH now automatically chunks massive PDFs and files into manageable semantic segments to bypass LLM context limits.

✨ Features

🧠 Dual Backend Support: seamless switching between OpenAI (Cloud) and Ollama (Local/Private).
🗣️ RAG Capabilities: "Chat" with your research topics using advanced Retrieval-Augmented Generation.
📚 Smart Storage: Automatic content deduplication and vector embedding using ChromaDB.
🔍 Semantic Search: Find what you need using natural language, not just keywords.
📂 Topic Organization: Keep different research streams (e.g., "AI Agents", "React Patterns") isolated and organized.
⚡ Fast & Efficient: Built on fastmcp and langchain for high performance.

📦 Installation

Recommended: using `uv` (Fastest)

# Run the AI Agent (MCP Server)
uvx kavi-research-assistant-mcp

# Run the Web UI (Gradio)
uv run kavi-research-ui

Using `pip`

pip install kavi-research-assistant-mcp

🎨 Web Interface (UI)

We provide a beautiful, colorful web interface to manage your research.

uv run kavi-research-ui

🎓 Ask Researcher: Chat with your research librarian.
💾 Save Knowledge: Easily paste and save new notes.
📊 Dashboard: View summaries and manage your topics.

⚙️ Configuration

You can configure the agent to use either OpenAI (default) or a local Ollama instance.

Option 1: OpenAI (Default)

Powerful, zero-setup (requires API Key).

export OPENAI_API_KEY=sk-...
export RESEARCH_DB_PATH=~/research_db
export LLM_PROVIDER=openai

Option 2: Ollama (Local & Private)

Run entirely on your machine. No API keys required.

Pull Models:

ollama pull llama3.2
ollama pull nomic-embed-text

Configure Environment:

export RESEARCH_DB_PATH=~/research_db
export LLM_PROVIDER=ollama
# Optional overrides
# export OLLAMA_BASE_URL=http://localhost:11434

Claude Desktop Setup

Add this to your claude_desktop_config.json:

{
  "mcpServers": {
    "kavi-research": {
      "command": "uvx",
      "args": ["kavi-research-assistant-mcp"],
      "env": {
        "RESEARCH_DB_PATH": "/Users/username/research_db",
        "OPENAI_API_KEY": "sk-..." 
      }
    }
  }
}

🛠️ MCP Tool Reference

Model Context Protocol (MCP) allows Kavi to act as a bridge between your AI and a private knowledge base. Below are the tools provided:

1. 📥 Data Ingestion

save_research_data(content: List[str], topic: str): Saves raw text or snippets.
- Usecase: Saving paper abstracts or news headlines.
save_research_files(file_paths: List[str], topic: str): Parses and vectorizes documents.
- Supported Formats: .pdf, .txt, .docx.
- Usecase: Ingesting a folder of research PDF papers.

2. 🔍 Knowledge Retrieval & RAG

ask_research_topic(query: str, topic: str): Answers questions using Retrieval Augmented Generation.
- Usecase: "What does my research say about Agentic Workflows?"
summarize_topic(topic: str): Generates a high-level executive summary of an entire library.
- Usecase: Periodic review of a project topic.

3. 📋 Management

list_research_topics(): Returns a list of all libraries and their document counts.
search_research_data(query: str, topic: str): Performs raw semantic similarity search for specific chunks.

🧪 Testing & Usage Steps

Step 1: Initialize the Environment

Ensure your preferred LLM backend is running. For Ollama:

ollama serve
ollama pull llama3.2
ollama pull nomic-embed-text

Step 2: Launch the Assistant

You can interact via the MCP Inspector (Command Line) or the Web UI.

To test via MCP Inspector:

npx @modelcontextprotocol/inspector uv run kavi-research-assistant-mcp

Once the inspector opens in your browser, you can manually trigger tools like list_research_topics.

Step 3: Populate with Knowledge

Ask your AI (via Claude Desktop or the UI) to save information:

"Save the following text to my 'ai-market' topic: [Your Text Here]"

Step 4: Validate RAG (The "Proof of Work")

Ask a question that only your saved data could answer:

"Based on my 'ai-market' data, what was the projected growth for 2026?"

Step 5: Dashboard Review

Open the UI to see your topic cards visualized gracefully.

uv run kavi-research-ui

💡 Typical Usecase Scenarios

Academic Research: Upload 50 PDF papers into a topic called thesis. Use ask_research_topic to find contradictions or common methodologies across all papers.
Market Intelligence: Save daily news snippets about competitors into competitor-intel. Every Friday, run summarize_topic to get a weekly briefing.
Code Library: Save documentation for obscure libraries into dev-docs. Use Kavi to answer "How do I implement X using Y?" without the LLM hallucinating.