Azure Analytics and AI Services for Architects: Synapse, Fabric, Databricks, Azure AI Foundry, OpenAI and Cognitive Services
Analytics Choices Recap
| Need | Service |
|---|---|
| SQL warehouse and BI | Synapse dedicated SQL or Fabric Warehouse |
| All-in-one SaaS analytics | Microsoft Fabric (OneLake, lakehouse, pipelines, Power BI) |
| Spark data engineering and ML | Azure Databricks or Fabric/Synapse Spark |
| Real-time analytics | Event Hubs + Stream Analytics or Fabric Real-Time Intelligence / Data Explorer |
| Ad-hoc SQL on lake files | Synapse serverless SQL or Fabric SQL endpoint |
AI Services Map
| Service | What it does |
|---|---|
| Azure AI Foundry | Platform to build, evaluate and deploy AI apps and agents with models |
| Azure OpenAI | OpenAI models (GPT family, embeddings, image models) with Azure security, regional deployment and content filters |
| Azure AI Search | Search service with keyword, vector and hybrid search. Core of RAG |
| Azure AI services | Prebuilt APIs: Vision, Speech, Language, Translator, Document Intelligence |
| Azure Machine Learning | Train, track, deploy and monitor custom ML models with MLOps |
| Azure AI Content Safety | Detect harmful content |
Names change Azure AI branding changes often (for example Cognitive Services became Azure AI services). Check current names when you read exam objectives.
RAG Architecture
Retrieval-augmented generation (RAG) lets a model answer using your own data instead of guessing.
- Ingest: documents in Blob Storage are split into chunks
- Embed: an embedding model turns chunks into vectors
- Index: store text and vectors in Azure AI Search
- Retrieve: for a user question, search for the best chunks (hybrid search)
- Generate: send question and chunks to the Azure OpenAI chat model to write the answer with citations
- Secure: Entra ID auth, managed identity between services, private endpoints, per-user document filters
python
from openai import AzureOpenAI
from azure.identity import DefaultAzureCredential, get_bearer_token_provider
token = get_bearer_token_provider(DefaultAzureCredential(), "https://cognitiveservices.azure.com/.default")
client = AzureOpenAI(azure_endpoint="https://oai-shop-482.openai.azure.com", azure_ad_token_provider=token, api_version="2024-10-21")
reply = client.chat.completions.create(model="gpt-4o-mini", # your deployment name
messages=[{"role": "system", "content": "Answer only from the context."},
{"role": "user", "content": "Context: Returns accepted within 30 days.\nQuestion: What is the return window?"}])
print(reply.choices[0].message.content)
API versions and model names change. Check the current Azure OpenAI documentation before using this sample.
Architect's Checklist for AI Workloads
- Data privacy: your prompts and data are not used to train the base models in Azure OpenAI (confirm in current terms)
- Networking: private endpoints and disabled public access
- Capacity: token-based quotas and provisioned throughput for steady loads. Plan for 429 retries
- Cost: model choice, caching and token limits
- Safety: content filters, grounding and human review for risky actions
- Governance: monitor with Azure Monitor and keep evaluation datasets
Choosing Quickly
| Scenario | Choice |
|---|---|
| Read invoices and extract fields | Document Intelligence |
| Transcribe calls | Speech service |
| Chatbot over company documents | Azure OpenAI + AI Search (RAG) |
| Custom forecasting model with MLOps | Azure Machine Learning |
| Company-wide lakehouse and dashboards | Fabric |
| Large Spark ETL with notebooks | Databricks |
Practice
- Sketch a RAG flow for HR policy documents
- List three security controls for an Azure OpenAI deployment
- Pick a service for five AI scenarios from the table
Conclusion
Use Fabric, Synapse or Databricks for analytics, Azure Machine Learning for custom models and AI Foundry, Azure OpenAI and AI Search for generative AI with RAG. Secure them with private networking, managed identities and content safety.
Quick Summary: analytics recap; AI map; RAG; checklist; choosing.

