Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Conversational Search for the Cloud Integration...

Vadim Klimov
October 27, 2026
120

Conversational Search for the Cloud Integration Capability

Event: SAP TechEd Berlin 2026
Date: October 27, 2026
Speaker: Vadim Klimov
Session: Conversational search for the Cloud Integration capability

Vadim Klimov

October 27, 2026

Transcript

23 slides
  1. Talk to Your Integration Platform Conversational Search for the Cloud

    Integration Capability Vadim Klimov SAP Integration Architect SAP TechEd Berlin October 27-29, 2026
  2. You want to ask your SAP Cloud Integration tenant about

    the integrations it hosts, in natural language. You want an AI assistant to help you analyze and govern what is in the tenant. Most of the data needed to find answers is already in the tenant. Bringing that data within reach of an AI assistant is the challenge. So which techniques and technologies bridge that gap?
  3. AI Assistant & SAP Cloud Integration Tenant Content | Mind

    the Gap AI assistant LLM ? SAP Cloud Integration tenant
  4. MCP | Key Components AI assistant LLM MCP tool calls

    MCP server API calls (fetch data) SAP Cloud Integration tenant
  5. AI assistants access tenant content in real time, in a

    standardized way. At scale, one query fans out into hundreds of API calls adding latency for the user and load on the tenant.
  6. RAG | Key Components AI assistant LLM Plugin 2. Query

    (similarity search) 1. Embed query Vector database Embedding model 3. Upsert 2. Embed chunks Ingestion pipeline 1. Fetch data SAP Cloud Integration tenant
  7. Semantic search scales to thousands of artifacts tenant content is

    indexed once, and queries are offloaded to a vector database. Chunks are embedded independently the search resolves similarity but not relation.
  8. SAP Cloud Integration Tenant Content | Zoom Out to Relationships

    Cloud Integration runtime profile HTTP receiver adapter Supports Configured for Integration package Has channel AEM sender connection Contains artifact Has step Has step Next step Content modifier step Integration flow Externalizes Next step Groovy script step Externalized parameter Has step Next step Write variable step Calls Uses credential OAuth2 Client Credentials Script collection Contains script Groovy script function Uses adapter Writes to Variable Has channel Next step HTTP receiver connection
  9. Minimum Viable Graph | Where to Begin • Start with

    a Labeled Property Graph (LPG). Graph model • Move to the Resource Description Framework (RDF) only if you need formal semantic reasoning or strict ontology standards. • Start with a few key designtime and runtime artifact types. • Add the remaining artifact types as needed. Knowledge model • Incorporate message processing data. • Address unstructured data sources such as documentation last.
  10. Multilayer Graph | Layers Platform and Landscape Integration Assets Integration

    Operations Documentation • Landscape • Artifacts • Deployments • Specifications • Tenant capabilities • Integration definitions • Message processing • Tests • Interface contracts • Data lineage • Standards and guidelines
  11. Summary MCP RAG Knowledge Graph Primary use case Targeted inspection

    of a scoped set of artifacts, retrieved on demand from the tenant Elastically scaled semantic search across all tenant artifacts, resolved outside the tenant Scalable analysis that traverses complex dependencies and lineage between artifacts, offloaded from the tenant Limitation Stateless calls, so widening analysis scope increases performance overhead on the tenant Self-contained chunks ranked by resemblance, with no way to follow the relationships between them Substantial implementation effort, driven by the complexity of modelling and ingesting tenant content Data freshness Live As fresh as the last ingestion into the vector database As fresh as the last ingestion into the graph database Implementation complexity Low Medium High

More Decks by Vadim Klimov