Azure's reference RAG app shows its work—and its price tag

What it does

This is Microsoft’s reference implementation for running Retrieval-Augmented Generation on Azure. It provides a web frontend and Python backend that let users ask multi-turn questions over private documents, using Azure AI Search to retrieve relevant chunks and Azure OpenAI—currently gpt-4.1-mini—to synthesize answers. The repo includes sample HR policy documents for a fictitious company so you can run it end-to-end immediately.

The interesting bit

Unlike many black-box demos, this one renders the full thought process and citations behind each answer, letting you debug why the LLM said what it said. It also doubles as a pricing calculator tutorial: the README itemizes every metered Azure service you will spin up, from Document Intelligence per-page fees to OpenAI per-token costs.

Key highlights

• Multi-turn chat interface that displays citations and reasoning traces for each response
• Supports many document formats and cloud-based data ingestion pipelines via Azure AI Search
• Optional multimodal reasoning over image-heavy documents using Azure AI Vision
• Optional speech input/output and Microsoft Entra ID integration for access control
• Performance tracing via Application Insights and persistent chat history via optional Cosmos DB

Caveats

• The README carries a prominent security notice that the template is built for demonstration, not production, and requires additional hardening before real use
• Cost estimation involves a long list of metered Azure services—OpenAI, Search, Document Intelligence, Container Registry, and more—with no predictable total
• Default deployment target shifted to Azure Container Apps as of late 2024, with Azure App Service remaining an optional path

Verdict

Grab this if you are an Azure shop looking for an official, feature-rich RAG reference to adapt. Skip it if you want a cheap, vendor-neutral, or production-ready chatbot out of the box.