RAG Demo with Azure MCP

This repository demonstrates a simple Retrieval-Augmented Generation (RAG) workflow using Model Context Protocol (MCP) for Azure search and Azure OpenAI for generation.

Summary

This demo application:

1. Spins up a simple SSE-based MCP search server (mcp_server.py) that wraps your document index.
2. Provides a CLI client script (demo.py) that:
   • Connects over SSE to the MCP server
   • Sends a user query
   • Retrieves the top N matching document chunks
   • Feeds those chunks as context to Azure OpenAI
   • Prints out a grounded, concise answer (or “I don’t know.”)

---

Getting Started

Prerequisites

• Python 3.8+
• Azure subscription with Azure AI Search and Azure OpenAI resources
• A working Azure AI Search index (text or hybrid with integrated vectorization configured)
• Environment variables

Setup

1. Clone the repo

   git clone https://github.com/juichiache/mcp-rag-demo
   
   

2. Install dependencies

pip install -r requirements.txt

Start MCP Server with environment variables loaded

In a new terminal, run mcp_server script

python mcp_server.py

Sample output from mcp_server.py:

Run RAG

python demo.py

Tech Stack Summary

1. Language & Runtime: Python 3.8+ with asyncio
2. Config: python-dotenv
3. MCP Framework: mcp.server.fastmcp + mcp.client.sse
4. Search: Azure Cognitive Search (azure-search-documents)
5. LLM: Azure OpenAI (openai SDK)
6. Transport: Server-Sent Events (SSE)
7. Utilities: logging, json, pprint

[ demo.py CLI ] | | SSE → “search” { query, top:3 } ↓ [ MCP FastMCP server (mcp_server.py) ] | | Azure Search query → get top‐N hits ↓ Stream back JSONRPC “content” chunks over SSE | [ demo.py assembles chunks ] | | Azure OpenAI ChatCompletion(prompt with context) ↓ Final answer printed to console

Python & Asyncio

Runs the whole thing as async code (both the demo CLI and the SSE‐based search server). Uses asyncio for concurrency and streaming.

Client side (demo.py)

mcp.client.sse.sse_client for opening an SSE (Server-Sent-Events) connection mcp.client.session.ClientSession to talk the MCP “search” tool

Server side (mcp_server.py)

mcp.server.fastmcp.FastMCP to expose an SSE endpoint and route “search” calls

Azure AI Search

azure.search.documents.aio.SearchClient to query a search index azure.search.documents.indexes.aio.SearchIndexClient to build or update that index Powers the retrieval of document “chunks” (PDF pages, etc.)

Azure OpenAI

openai.AzureOpenAI client against your Azure OpenAI deployment (GPT-4.1) Takes the top N retrieved chunks, concatenates their "content" fields, and feeds them into a chat‐completion prompt

Server-Sent Events (SSE)

A lightweight streaming over HTTP mechanism so the client can send a prompt and incrementally receive back search hits

dotenv for configuration

python-dotenv to pull your Azure Search service name/key, Azure OpenAI endpoint/key, index name, etc. out of a .env file

Standard Python tooling

• logging for debug output
• pprint / json for introspecting responses
• asyncio.run(...) to kick off the async CLI