Embeddings and external models in Azure SQL Database

How the engine itself produces a vector, and how it is allowed to reach a model. Verified 2026-08-28 against both a live Azure SQL Database (General Purpose serverless, compatibility level 170) and the local container, using a real Azure OpenAI deployment created for the run. Option lists and limits were re-sourced from Microsoft Learn on 2026-09-03.

AI_GENERATE_CHUNKS splits text and CREATE EXTERNAL MODEL registers an endpoint; neither touches the network. AI_GENERATE_EMBEDDINGS and sp_invoke_external_rest_endpoint, the call it builds on, both do, and that is where everything below goes wrong. All four work in the container as well as in the cloud: the 2026 catalog note calling this surface cloud only was wrong, and a real embedding call completed from inside a container.

The correction