Agentic AI Assistant on SAP BTP

The Challenge

Enterprise AI systems built on fixed RAG pipelines can only answer questions in one way: retrieve context, generate response. Real business queries are more complex: they require deciding what information to look up, when to call a live system API, and how to combine multiple results before forming an answer. The goal of this PoC is to build a system on SAP BTP where the AI itself reasons about what action to take next, using available tools autonomously without hardcoding the logic for every query type. This demonstrates the agentic pattern the enterprise AI market is converging on.

Technology

CAP Service

LC

LangChain Agent

SAP AI Core

HANA Cloud

Architecture

My Role

• Designing the ReAct agent loop using LangChain with SAP AI Core as the underlying LLM provider
• Implementing HANA Cloud vector store as a retrieval tool available to the agent
• Exposing SAP OData endpoints as structured data tools the agent can call autonomously
• Building the CAP service layer as the API surface and audit trail for all agent interactions
• Instrumenting reasoning traces for observability, capturing every tool call and decision step

Outcomes

• Validates LangChain + SAP AI Core as a reusable agentic pattern on BTP
• Demonstrates multi-step reasoning with tool selection, going beyond standard RAG pipelines
• Establishes a foundation for more complex multi-agent architectures on SAP BTP