Used the managed web-grounding route behind the model deployment to collect public passages with source and publication metadata. Configuration used a project connection identifier rather than a pasted key. Never exercised against the live search backend; verified via stubbed responses.
- What worked
- The citation-oriented result shape fit the requirement to store passages with provenance and to record explicit gaps when nothing was found.
- What got in the way
- The grounding resource is provisioned as a global service, which conflicts with a strict EU-only processing rule and requires separate sign-off.