Edit

Share via

Select Azure PaaS solutions for AI

This article describes how to choose appropriate resources for Azure AI platform as a service (PaaS) solutions. The following table provides an overview of the primary Azure AI PaaS solutions and important decision criteria.

AI services AI type Description Skills required
Azure AI Foundry Generative AI and nongenerative AI A platform for building and deploying generative and nongenerative AI applications Developer and data science skills
Azure AI services Generative AI and nongenerative AI Various services that provide prebuilt generative and nongenerative AI models Developer skills
Azure OpenAI Generative AI A service for accessing OpenAI models Developer and data science skills
Azure Machine Learning Machine learning A service for training and deploying machine learning models Developer skills and advanced data science skills

Select resources for generative AI workloads

Generative AI combines various resources to process input data and generate meaningful outputs. Select the right resources to ensure that applications, such as applications that use retrieval-augmented generation (RAG), deliver accurate results by grounding AI models effectively.

Diagram that shows the basic components of a generative AI workload.

Generative AI workflow

The following workflow corresponds to the preceding diagram:

  1. The AI app receives a query from the user.

  2. An orchestrator, such as Azure AI Foundry Agent Service, Semantic Kernel, or LangChain, manages the data flow.

  3. A search and retrieval mechanism identifies the appropriate grounding data.

  4. The mechanism sends the grounding data to a generative AI platform.

  5. The generative AI platform generates a response based on the user query and grounding data.

Generative AI resource selection

Use the following recommendations to build generative RAG workloads:

  1. Choose a generative AI platform. Use Azure AI Foundry or Azure OpenAI to deploy and manage generative AI models. Azure AI Foundry provides a code-first platform that includes built-in tools to develop, deploy, and orchestrate applications. Use Azure OpenAI if you only need access to OpenAI models.

  2. Choose the appropriate AI compute type. Azure AI Foundry requires compute instances for specific capabilities. Select a compute type that meets your performance and budget needs.

  3. Pick an orchestrator. Use popular orchestrators like Azure AI Foundry Agent Service, Semantic Kernel, or LangChain to manage data flow and interactions. For workloads that have multiple collaborating agents, your orchestrator must support the AI agent orchestration patterns that you use.

  4. Pick a search and knowledge retrieval mechanism. To ground generative AI models, create an index or vector database for relevant data retrieval. Use Azure AI Search to build traditional and vector indexes from various data sources, apply data chunking, and use multiple query types. For structured databases, consider Azure Cosmos DB, Azure Database for PostgreSQL, or Azure Cache for Redis.

  5. Choose a data source for grounding data. Store grounding data in Azure Blob Storage for images, audio, video, or large datasets. Alternatively, use databases that AI Search or vector databases support.

  6. Pick a compute platform. Use the Azure compute decision tree to select the right platform for your workload.

Select resources for nongenerative AI workloads

Nongenerative AI workloads use platforms, compute resources, data sources, and data processing tools to support machine learning tasks. Select the right resources to ensure that you can build AI workloads by using prebuilt or custom solutions effectively.

Diagram that shows the basic components of a nongenerative AI workload.

Nongenerative AI workflow

The following workflow corresponds to the preceding diagram:

  1. The AI app ingests incoming data.

  2. An optional data processing mechanism extracts or manipulates the data.

  3. An AI model endpoint analyzes the data.

  4. Data can be used for training or fine-tuning of AI models.

Nongenerative AI resource selection

Use the following recommendations to build nongenerative AI workloads:

  1. Choose a nongenerative AI platform. Use AI services or Machine Learning based on your requirements. AI services provides prebuilt AI models that simplify deployment and reduce the need for deep data science expertise. Machine Learning provides a platform to develop custom machine learning models by using your data. It also integrates those models into your workloads.

  2. Choose the appropriate AI compute type. Machine Learning requires compute resources to run jobs or host endpoints. Select a compute type that meets your performance and budget needs. AI services don't require compute resources.

  3. Pick a data source. Use supported data sources to host training data for Machine Learning. Many AI services don't require fine-tuning data. And some AI services, like Azure AI Custom Vision, allow you to upload local files to a managed data storage solution.

  4. Pick a compute platform. Use the Azure compute decision tree to select the right platform for your workload.

  5. Pick a data processing service (optional). Use Azure Functions to process serverless data or Azure Event Grid to trigger data processing pipelines.

Next step

Networking