Concept card

SAP Prompt Registry

As of 2026-09-27

Edited by Cédric Mary, sole editor and publication director — 27 years in the SAP ecosystem. How these pages are produced, and where automation is used.

What is SAP Prompt Registry?

The prompt registry of SAP AI Core (generative AI hub) stores prompt templates and orchestration configs as named, SemVer-versioned assets, either imperatively through CRUD APIs or declaratively from Git. It gives history, reuse and runtime referencing — but no built-in approval workflow: review gates come from your Git process.

What it is

The prompt registry is a component of the generative AI hub in SAP AI Core that "integrates prompt templates into SAP AI Core, making them discoverable across your applications and orchestration". It manages the lifecycle of two asset types: prompt templates and, since February 2026, orchestration configs — the full description of an orchestration call (model, parameters, filters, masking, grounding). Both can be consumed directly through the API or referenced from orchestration requests. The registry is reached through the AI API (/v2/lm/promptTemplates and related endpoints), through AI Launchpad and through the SAP Cloud SDK for AI in JavaScript, Java and Python.

Anatomy of a prompt template

A template has a name, a version, a scenario and a spec. The spec holds the message template — system, user and assistant roles with placeholders written {{?placeholder}} — optional defaults for placeholders and an unstructured additional_fields object for metadata such as model parameters or a list of blocked models. Versions must follow semantic versioning; the latest iteration is the head version, flagged isVersionHead: true, and every iteration is visible through the history endpoint. Templates live at tenant level by default; adding the headers AI-Resource-Group-Scope: true and AI-Resource-Group scopes them to one resource group, which is how you separate teams or environments. At runtime you fill a template by ID or by the triple scenario, name and version through the substitution endpoint, or reference it inside an orchestration call with template_ref, by ID or by name, scenario and version, optionally with scope resource_group or tenant.

Why it matters

Without a registry, prompts live in code and environment variables where nobody can reconstruct what the system was told on a given date.
The registry has no approval workflow of its own — teams that assume one ship unreviewed prompts; Git-based declarative mode is where review becomes enforceable.
Template history alone is not evidence: the version used by each inference must be logged by the application or by inference observability.

Key points

Component of the generative AI hub: manages prompt templates and (since Feb 2026) orchestration configs as reusable, discoverable assets.
Template = name + version (SemVer) + scenario + spec (messages with {{?placeholders}}, defaults, additional_fields).
Head version flagged isVersionHead: true; history endpoint lists every iteration.
Scope: tenant by default; resource-group scope via AI-Resource-Group-Scope: true + AI-Resource-Group headers.
Imperative API (CRUD, design time) vs declarative (Git files *.prompttemplate.ai.sap.yaml, synced, managedBy: declarative, CI/CD).
Runtime use: substitution endpoint by ID or scenario/name/version, or template_ref inside an orchestration call.
No built-in reviewer role or approval gate: enforce review through protected branches and pull requests in declarative mode.
Prompt optimization writes improved prompts back to the registry; evaluations compare templates and models on your dataset.

Terms used on this page

Prompt template: Versioned, reusable prompt definition with role-based messages, placeholders, defaults and metadata, stored in the prompt registry.
Head version: The latest iteration of a template, flagged isVersionHead: true.
Imperative API: CRUD interface of the registry, recommended for design-time iteration; changes are tracked in the history endpoint.
Declarative API: Git-based management of templates and configs, synchronised by AI Core; recommended for runtime and CI/CD.
template_ref: Reference to a stored template inside an orchestration request, by ID or by name, scenario and version.
Substitution: Endpoint that fills a template's placeholders with input values and returns the resulting messages.
Orchestration config (stored): Versioned description of an orchestration call kept in the registry and referenced from requests.
Prompt optimization: Hub feature that improves a registry template against a dataset and metric and saves the result back to the registry.

Sources

Full card available to members. What the full card adds: the full decision framework · the SAP vs Snowflake / Databricks / Fabric comparison · the common pitfalls and their fix · the cheat sheet · the architecture schemas · the code blocks · the facts worth quoting.

Open in the app →