AI Extension - Foundry Project, APIM integration and Usage Metrics

[nkidambi]

What service are you using?

Azure Database for PostgreSQL

What feature or capability is this about?

azure_ai_extensions, APIM, GPT 5 series support

What's your feedback?

I currently have a knowledge agent that does SQL, Hybrdi and Graph based retrieval using PG AGE v18. Some of my customers are using it. I notice the following constraints/shortcomings when deploying the solution beyond POC.

• How do get the usage metrics for the AI calls made by PG? s it possible to point the endpoint value to an APIM resource with Foundry OpenAI endpoint as backend?
  APIM has extensive token metrics tracking. But using PG extension that does not leverage that does not give the accurate token consumption and estimating cost for our customers becomes difficult.

• Use Foundry project URL instead of Foundry OpenaI URL
  Ability to use Project endpoints, enables us to use agent frameworks ResponsClient across the board. Also mitigates additional settings to be configured when deployed.

• ai_generate extension truncates response and not able to set max_tokens are parameter
  Ability to set max_tokens helps to provide richer response when using ai_generate. Currently, the extension is truncating and the response is not rich enough for hetergenous data

• Does AI Etensions support GPT 5 series
  With GPT 4.1 model being retired, would GPT 5 be supported in ai_generate along with the ability to speicify the reasoning effort?

[Shreya Aithal (shreyaaithal)]

Thank you for the valuable feedback, nkidambi - really appreciate you taking the time to document these so thoroughly.

APIM integration, Foundry Project URL support, and GPT-5 series are already available in Azure HorizonDB and will be coming to Azure Database for PostgreSQL soon. The max_tokens parameter limitation is acknowledged and being picked up on priority.

We'll update this thread with confirmed ETAs, likely in a couple of weeks.

[nkidambi]

Shreya Aithal (@shreyaaithal) Thanks for the update.
One more thing... The APIM integration is crucial for production release as it helps my customers to load balance the Foundry. Without that, providing the AOAI endpoint directly in PG extensions, completely disables the load balancing capability.