Azure AI Services

Microsoft's AI stack on Azure spans three layers: Azure OpenAI Service for direct access to OpenAI models, Azure AI Foundry for multi-model generative AI with agents and evaluation, and Azure Machine Learning for classical ML training and MLOps. Task-specific services (Document Intelligence, AI Search, AI Vision, AI Language) round out the platform.

Generative AI

Azure OpenAI Service GPT-4o, o1, o3-mini, DALL·E 3, Whisper — with Entra ID auth, content filtering, and On-Your-Data RAG.

Azure AI Foundry Multi-model catalog, Agent Service, Prompt Flow, evaluation, and safety in one workspace.

Azure Machine Learning Training, AutoML, hyperparameter sweeps, managed endpoints, MLOps pipelines.

Retrieval & Document AI

Azure AI Search Hybrid (BM25 + vector) search with semantic reranking — the default RAG retriever on Azure.

AI Document Intelligence OCR, forms, tables, invoices, IDs, contracts — Markdown output for LLM ingestion.

AI Vision & Language Image analysis, multimodal embeddings, sentiment, NER, PII redaction, summarization.

Choosing Between Azure AI Services

Just need GPT-4o / o1 / embeddings? → Azure OpenAI Service.
Building a generative app with agents, evaluation, multiple models? → Azure AI Foundry.
Training a custom classical model (XGBoost, PyTorch) or running MLOps pipelines? → Azure Machine Learning.
Extracting structured fields from PDFs / forms? → Document Intelligence.
RAG retrieval over enterprise documents? → Azure AI Search (with Document Intelligence for parsing).
Well-defined NLP task (sentiment, NER, PII)? → Azure AI Language — faster and cheaper than GPT.
Image understanding or multimodal search? → Azure AI Vision.

Azure vs. AWS Equivalents

Azure	AWS Equivalent	Notes
Azure OpenAI Service	Amazon Bedrock	Azure hosts OpenAI; Bedrock hosts Anthropic/Meta/Mistral/Cohere/Amazon (not OpenAI).
Azure AI Foundry	Bedrock + SageMaker Studio	Foundry unifies model catalog + agents + evaluation.
Azure Machine Learning	Amazon SageMaker	Closest peers — full ML platform.
Azure AI Search	Amazon OpenSearch + Kendra	AI Search's semantic reranker has no direct AWS equivalent.
AI Document Intelligence	Amazon Textract	Similar features; Document Intelligence emits Markdown natively.
Azure AI Vision	Amazon Rekognition	Florence-backed; Rekognition stronger on video/faces.
Azure AI Language	Amazon Comprehend	Similar NLP primitives; Azure bundles QnA + CLU.

How the services fit together

The services form layers rather than alternatives. Azure OpenAI Service and the Foundry model catalog supply models. Azure Machine Learning handles training and deployment of your own models. Task-specific services (Document Intelligence, Vision, Language) solve narrow problems with pretrained models and predictable cost. Azure AI Search provides retrieval. Azure AI Foundry ties these together for generative applications with agents, prompt flows, evaluation and safety controls, and all of them share Microsoft Entra ID authentication, private networking and Azure Monitor logging.

A typical RAG stack on Azure

A common enterprise pattern combines four services. Document Intelligence's Layout model converts PDFs and scans into Markdown with tables preserved. The text is chunked, embedded with an Azure OpenAI embedding model, and indexed in Azure AI Search, which combines keyword (BM25) and vector search with optional semantic reranking. At question time the application retrieves the top passages and sends them with the question to a chat model, and Foundry evaluations measure groundedness and relevance over a test set. Content filters and managed identities apply at each step.

All pages in this section

Azure OpenAI ServiceazureOpenAI models with Azure security, quotas and content filtering. Azure AI Foundryazuremodel catalog, agents, prompt flow and evaluation. Azure Machine Learningazuretraining, AutoML, endpoints and MLOps. Azure AI Searchazurehybrid search with semantic reranking for RAG. Azure AI Document IntelligenceazureOCR, layout and field extraction. Azure AI Vision & Languageazureimage analysis and NLP task APIs.