Generative AI Hub in SAP AI Core: Models, Orchestration, Prompt Registry and Content Filtering

Keep this: job aid

Generative AI hub design card

• Extended plan required for the hub; 30-day basic trial for learning; resource group per use case (max 50 per tenant).
• Scenarios: foundation-models (one deployment per model) or orchestration (one deployment, model in config).
• Claude via aws-bedrock; SAP-ABAP-1 via orchestration only; check SAP Note 3437766 for models, rates, limits, deprecation.
• Pinned version stops on its deprecation date; latest auto-upgrades. Re-run the regression set before patching.
• V2: POST /v2/completion; headers Authorization and AI-Resource-Group; body config.modules + placeholder_values.
• Modules: templating + model (mandatory); grounding, masking, filtering, translation (optional); fallbacks on region, 429, 408, 5xx.
• Masking: anonymize (irreversible) or pseudonymize (restored); detection not guaranteed, so minimize first.
• Filtering: Azure Content Safety (0/2/4/6, Prompt Shield) and Llama Guard 3; each filter type billed separately.
• Prompt registry: imperative API for design, declarative Git sync for CI/CD; version, review, regression-test every change.
• Cost: tokens → GenAI tokens → capacity units; output tokens cost more; measure tokens per request in test.