Microsoft Azure Virtual Training Day

​

Develop Your Own Custom Copilots with Azure AI

1. Introduction to Azure AI Studio

• Central platform for building custom copilots & generative AI apps.

• Features: pro-code development, prompt/model orchestration, fine-tuning, evaluations, secure connections, deployment endpoints.

• Environment setup with AI hubs & projects, connections to Azure or external APIs, and RBAC role-based governance.

2. Responsible AI & Governance

• Microsoft’s 4-stage Responsible AI plan: Identify harms → Measure → Mitigate → Operate responsibly.

• Tools: Azure AI Content Safety (prompt shields, groundedness detection, protected material detection, custom harm categories).

• Collaboration roles via RBAC (Owner, Contributor, Reader, AI Developer, Inference Deployment Operator, Custom roles).

3. Model Catalog & Deployment

• Models available: BERT, GPT, LLaMA, Phi-3-mini, plus Azure OpenAI offerings.

• Options: deploy to endpoints (serverless or managed compute), fine-tune, or test in playground.

• Use benchmarks (accuracy, coherence, fluency, relevance, groundedness) to evaluate models before deployment.

4. Model Optimization Strategies

• Prompt engineering → refine instructions/system messages for better responses.

• RAG (Retrieval-Augmented Generation) → ground copilots in your own data using Azure AI Search + embeddings.

• Fine-tuning → domain-specific adaptation with custom JSONL training datasets.

• Combined strategies maximize contextual accuracy, style consistency, and reliability.

5. Prompt Flow & Custom Copilots

• Lifecycle: Initialization → Experimentation → Evaluation → Production.

• Flow Components: Inputs, Nodes (tools), Outputs, prompt/LLM/Python tools.

• Variants: experiment with different prompts, system messages, or models.

• Build custom copilots with RAG + Prompt Flow → integrate your own indexed data for grounded answers.

6. Evaluation & Code-First Development

• Evaluate copilots with benchmarks, manual ratings, AI-assisted metrics (accuracy, coherence, fluency, safety).

• Built-in metrics via Prompt Flow evaluations.

• Code-first dev tools: Azure AI SDKs, Jupyter, VS Code, Semantic Kernel, LangChain, Cognitive Search, OpenAI Service.​​​​​​​​​​​​​​​​