GoomeGum/mcp-server

MCP Server - Context Enhancement Tool

A FastAPI-based Model Context Protocol (MCP) server that enhances simple text contexts into elaborate, detailed prompts with caching functionality. Developed for Google Colab with GPU support.

Features

• Context Elaboration Tool: Transform simple text contexts into detailed, elaborate prompts
• Cached Prompts: Access pre-generated elaborate prompts by concept ID
• Prompt Cache Management: List all cached elaborate prompts
• CSV Data Loading: Load pre-trained elaborate prompts from CSV files

Project Structure

mcp_server/
├── main.py                      # FastAPI application with MCP endpoints
├── models.py                    # Pydantic models for request/response schemas
├── cache.py                     # Caching functionality and prompt elaboration
├── requirements.txt             # Python dependencies
├── similarity_results_train.csv # Training data for cached elaborate prompts
├── MCPServer.ipynb              # Google Colab notebook implementation
└── README.md                    # This file

API Endpoints

Tools

• POST /tool/get_elaborate_description_prompt - Transform simple context into elaborate prompts

Resources

• GET /resource/cached_description/{concept_id} - Get cached elaborate prompt by ID
• GET /resource/cached_description_list - List all cached elaborate prompts

Quick Start

For Google Colab (Recommended)

1. Open the MCPServer.ipynb notebook in Google Colab
2. Run all cells to install dependencies and start the server
3. The server will create a public ngrok tunnel - access the API at: {ngrok_url}/docs

For Local Development (GPU is needed)

1. Install dependencies:

   pip install -r requirements.txt
   

2. Access the API documentation at: http://localhost:8000/docs

Usage Examples

Generate Elaborate Prompt

curl -X POST "http://localhost:8000/tool/get_elaborate_description_prompt" \
     -H "Content-Type: application/json" \
     -d '{"context": "a sunset over mountains"}'

Get Cached Elaborate Prompt

curl "http://localhost:8000/resource/cached_description/concept123"

List All Cached Elaborate Prompts

curl "http://localhost:8000/resource/cached_description_list"

Data Format

The server loads cached elaborate prompts from similarity_results_train.csv with the following format:

• input: The simple concept/context identifier
• finetuned_model_answer: The elaborate, enhanced prompt

Development

This server is designed to work with Phi-4 or similar language models for transforming simple text contexts into elaborate, detailed prompts. The project is optimized for Google Colab with GPU support for running the fine-tuned model efficiently.

Key Features for Colab:

• ngrok integration: Creates public tunnels for accessing the server externally
• GPU support: Leverages Colab's free GPU for model inference
• Pre-configured setup: All dependencies and model loading handled in the notebook
• Easy access: Server documentation available at {ngrok_url}/docs after running

Local Development Notes:

The current implementation includes placeholder logic in cache.py that can be replaced with actual model calls for context enhancement and prompt elaboration when running locally without GPU access.