Skip to content

🎯 Orchestrator

The Orchestrator is the core engine of GPT-RAG, an agentic orchestration layer built on the Microsoft Agent Framework and Azure AI Foundry Agent Service. It coordinates agent-based RAG workflows — each agent has a defined role — to generate accurate, context-aware responses for complex user queries. GitHub Repository.

Key Features

  • Strategy-Based Architecture: Pluggable orchestration strategies selected via Azure App Configuration (AGENT_STRATEGY).
  • Context Retrieval: Intelligent retrieval from Azure AI Search with citation support.
  • Microsoft Agent Framework: Built on the Microsoft Agent Framework.
  • Conversation Persistence: Maintains conversation history in Cosmos DB with real-time SSE streaming.
  • Extensible Design: Easy to add new strategies by extending BaseAgentStrategy.

Available Strategies

The Orchestrator supports multiple strategies. The active strategy is set via the AGENT_STRATEGY key in Azure App Configuration. The default is maf_lite.

Key Strategy Description
maf_lite MAF Lite (default) Microsoft Agent Framework with direct Azure OpenAI model access. Lightweight — no Agent Service dependency. Includes user profile memory and optional agentic search.
maf_agent_service MAF + Agent Service Microsoft Agent Framework with Azure AI Foundry Agent Service for server-side thread management and tool orchestration. Includes user profile memory and optional agentic search.
single_agent_rag Single Agent RAG Uses Azure AI Agents SDK with Agent Service for agentic RAG. Supports dynamic routing, streaming via event handlers, and pre-warming for low-latency first responses.
mcp MCP Model Context Protocol strategy using Semantic Kernel. Connects to an MCP server for tool orchestration and passes user context via HTTP headers.
nl2sql NL2SQL Natural language to SQL translation using a multi-agent pipeline (Triage → SQL Query → Synthesizer) with Semantic Kernel and Agent Service.

Visual Guide

New to the Orchestrator? Check out the Orchestrator Visual Guide for a visual walkthrough of the architecture and key components.

Repository

🔗 GitHub Repository

© 2025 GPT-RAG — powered by ❤️ and coffee ☕