AI Apps
Azure OpenAI Service
The language, reasoning, and embedding models a workload calls, deployed inside the customer's own Azure subscription and network.
Product
Azure OpenAI ServiceAzure OpenAI Service provides the language, reasoning, embedding, and realtime audio models an AI workload calls, deployed into the customer's own Azure subscription. Prompts and responses are not used to train the models. Deployments are regional, so data residency is a configuration rather than a promise, and access runs through Microsoft Entra ID and network controls like any other Azure resource. Function calling lets a model request a tool the application defines, embeddings power the vector side of retrieval, and the realtime models carry speech in and speech out on one connection instead of a transcription and synthesis chain.
How it works
- A deployment is created in a chosen region with an agreed quota, so latency and data residency are decided rather than inherited.
- The application authenticates with a managed identity through Microsoft Entra ID, so no model key is stored in application configuration.
- Embedding models turn documents and questions into vectors, which is what makes hybrid retrieval on Azure AI Search possible.
- Chat and reasoning models receive the retrieved context as grounding, with the prompt assembled by the application rather than by the model.
- Function calling lets the model request one of the tools the application has declared, and the application decides whether to run it.
- Realtime models carry a voice conversation on a single connection, avoiding the delay a separate transcription and synthesis chain adds.
- Content filters, quota, and per deployment metrics are monitored, so cost and refusal rates are visible before they become a support ticket.
A problem it solves
A team wants to use current language models but cannot send corporate documents to a public endpoint, and cannot tell a regulator where the data was processed. Azure OpenAI Service puts the deployment inside the customer's own subscription and region, behind their identity and network controls, so the question of where the data went has an answer in the resource configuration.