Build and maintain your company Copilot with Azure ML and GPT-4
summary
Demo builds a Contoso customer-service Copilot that uses prompt flow + LangChain to run a three-step RAG pipeline: intent classification, Cognitive Search retrieval, then GPT-4 generation, all runnable in bulk tests and deployable with one click to an Azure ML endpoint. Iteration shows accuracy rising from ~61% to ~72% with prompt variants and groundedness scoring of 4.31/5 using GPT-4 evaluators. Supports vector indices (Faiss or Cognitive Search), fine-tuning via the model catalog, live monitoring, and automated experiment comparisons.
AI-written summary · about
chapters (13)
- 0:00Introduction to AI on Azure
- 1:30AI adoption across industries
- 3:35The Azure AI services stack
- 6:11Understanding RAG
- 10:31Customer service co-pilot demo
- 13:58Building complex AI workflows
- 20:04Introduction to prompt flow
- 24:23Bulk testing and metrics
- 28:11Iterating prompts with variants
- 30:56Evaluating RAG groundedness
- 37:14Deploying and monitoring
- 40:17Model catalog and fine-tuning
- 42:29Conclusion and resources
related talks
01h 58mBuild, Evaluate and Deploy a RAG-Based Retail Copilot with Azure AI
45mBuilding a copilot: Azure AI Studio or Copilot Studio
46mBuilding Copilots - key lessons and best practices
27mBuild a RAG-based Retail Copilot Code-First on Azure AI Foundry
39mDisrupt and Grow with Microsoft Azure AI
10mPrompting your next company strategy
36mCustom copilot Teams Integration in Government applications
36mWhat's new with Copilot Studio