Sap AI Core

SAP AI Core & AI Launchpad Skill

Related Skills

• sap-btp-cloud-platform: Use for platform context, BTP account setup, and service integration
• sap-cap-capire: Use for building AI-powered applications with CAP or integrating AI services
• sap-cloud-sdk-ai: Use for SDK integration, AI service calls, and Java/JavaScript implementations
• sap-btp-best-practices: Use for production deployment patterns and AI governance guidelines

When to Use This Skill

Use this skill when provisioning SAP AI Core, using SAP AI Launchpad, configuring Generative AI Hub orchestration, choosing model providers, building RAG or grounding flows, managing prompt templates, deploying training/inference workloads, or wiring AI capabilities into SAP applications.

Table of Contents

1. Overview
2. Quick Start
3. Service Plans
4. Model Providers
5. Orchestration
6. Content Filtering
7. Data Masking
8. Grounding (RAG)
9. Tool Calling
10. Structured Output
11. Embeddings
12. ML Training
13. Deployments
14. Bundled Resources
15. SAP AI Launchpad
16. Prompt Registry
17. API Reference
18. Common Patterns
19. Troubleshooting
20. References

Overview

SAP AI Core is a service on SAP Business Technology Platform (BTP) that manages AI asset execution in a standardized, scalable, hyperscaler-agnostic manner. SAP AI Launchpad provides the management UI for AI runtimes including the Generative AI Hub.

Core Capabilities

Three Components

1. SAP AI Core: Execution engine for AI workflows and model serving
2. SAP AI Launchpad: Management UI for AI runtimes and GenAI Hub
3. AI API: Standardized lifecycle management across runtimes

Quick Start

Prerequisites

• SAP BTP enterprise account
• SAP AI Core service instance (Extended plan for GenAI)
• Service key with credentials

1. Get Authentication Token

# Set environment variables from service key
export AI_API_URL="<your-ai-api-url>"
export AUTH_URL="<your-auth-url>"
export CLIENT_ID="<your-client-id>"
export CLIENT_SECRET="<your-client-secret>"

# Get OAuth token
AUTH_TOKEN=$(curl -s -X POST "$AUTH_URL/oauth/token" \
  -H "Content-Type: application/x-www-form-urlencoded" \
  -d "grant_type=client_credentials&client_id=$CLIENT_ID&client_secret=$CLIENT_SECRET" \
  | jq -r '.access_token')

2. Create Orchestration Deployment

# Check for existing orchestration deployment
curl -X GET "$AI_API_URL/v2/lm/deployments" \
  -H "Authorization: Bearer $AUTH_TOKEN" \
  -H "AI-Resource-Group: default" \
  -H "Content-Type: application/json"

# Create orchestration deployment if needed
curl -X POST "$AI_API_URL/v2/lm/deployments" \
  -H "Authorization: Bearer $AUTH_TOKEN" \
  -H "AI-Resource-Group: default" \
  -H "Content-Type: application/json" \
  -d '{
    "configurationId": "<orchestration-config-id>"
  }'

3. Use Harmonized API for Model Inference

ORCHESTRATION_URL="<deployment-url>"

curl -X POST "$ORCHESTRATION_URL/v2/completion" \
  -H "Authorization: Bearer $AUTH_TOKEN" \
  -H "AI-Resource-Group: default" \
  -H "Content-Type: application/json" \
  -d '{
    "config": {
      "module_configurations": {
        "llm_module_config": {
          "model_name": "gpt-4o",
          "model_version": "latest",
          "model_params": {
            "max_tokens": 1000,
            "temperature": 0.7
          }
        },
        "templating_module_config": {
          "template": [
            {"role": "system", "content": "You are a helpful assistant."},
            {"role": "user", "content": "{{?user_query}}"}
          ]
        }
      }
    },
    "input_params": {
      "user_query": "What is SAP AI Core?"
    }
  }'

Service Plans

Key Restrictions:

• Free and Standard mutually exclusive in same subaccount
• Free → Standard upgrade possible; downgrade not supported
• Max 50 resource groups per tenant

Model Providers

SAP AI Core provides access to model providers through a tenant-specific catalog. Treat exact model names and versions as examples until verified in the target tenant with GET /v2/lm/scenarios/foundation-models/models or SAP AI Launchpad Model Library.

• Azure OpenAI: GPT-family chat, vision, reasoning, realtime, and embedding models where entitled
• SAP Open Source: Llama/Falcon/Mistral-family open source models where enabled
• Google Vertex AI: Gemini-family chat, vision, code, and embedding models where entitled
• AWS Bedrock: Anthropic Claude and Amazon model families where entitled
• Mistral AI: Mistral Large/Small/Codestral-family models where enabled
• IBM: Granite models
• Perplexity: Sonar-family web-grounded models where enabled

For detailed provider configurations and model lists, see references/model-providers.md.

Orchestration

The orchestration service provides unified access to multiple models through a modular pipeline with 8 execution stages:

1. Grounding → 2. Templating (mandatory) → 3. Input Translation → 4. Data Masking → 5. Input Filtering → 6. Model Configuration (mandatory) → 7. Output Filtering → 8. Output Translation

For complete orchestration module configurations, examples, and advanced patterns, see references/orchestration-modules.md.

Content Filtering

Azure Content Safety: Filters content across 4 categories (Hate, Violence, Sexual, SelfHarm) with severity levels 0-6. Azure OpenAI blocks severity 4+ automatically. Additional features include PromptShield and Protected Material detection.

Llama Guard 3: Covers 14 categories including violent crimes, privacy violations, and code interpreter abuse.

Data Masking

Two PII protection methods:

• Anonymization: MASKED_ENTITY (non-reversible)
• Pseudonymization: MASKED_ENTITY_ID (reversible)

Supported entities (25 total): Personal data, IDs, financial information, SAP-specific IDs, and sensitive attributes. For complete entity list and implementation details, see references/orchestration-modules.md.

Grounding (RAG)

Integrate external data from SharePoint, S3, SFTP, SAP Build Work Zone, and DMS. Supports PDF, HTML, DOCX, images, and more. Limit: 2,000 documents per pipeline with daily refresh. For detailed setup, see references/grounding-rag.md.

Tool Calling

Enable LLMs to execute functions through a 5-step workflow: define tools → receive tool_calls → execute functions → return results → LLM incorporates responses. Templates available in templates/tool-definition.json.

Structured Output

Force model responses to match JSON schemas using strict validation. Useful for structured data extraction and API responses.