nsantra/rag-mcp-server ? reverse-engineered prompt

Build me a Python MCP server for a RAG workflow that I can connect to Claude Desktop and Cursor.

I want it to manage a persistent Chroma vector database with separate collections for different knowledge areas, let me add PDFs from a local file or a URL, and then retrieve the most relevant chunks with citations when I ask questions. Please include support for local Hugging Face embeddings and reranking, with an option to use GPU or CPU depending on my setup.

It should also let me create, list, rename, inspect, and delete collections and documents, and have a way to show all available tools automatically. Please make the setup simple with an env file for paths and model settings, and include clear defaults for local model folders. If it helps, look up current docs online for MCP, FastMCP, ChromaDB, and LangChain while you build it.