Hybrid Grounding on HANA Cloud — Vector Engine + Knowledge Graph Engine

Why one grounding technique is not enough

SAP's own Knowledge Graph lesson on SAP Learning lists the limits of naive vector RAG: it overlooks relationships between entities, and LLMs struggle with enterprise formats, hierarchies and domain knowledge. The remedy SAP names is to combine GraphRAG with vector-based RAG over business data in SAP HANA and Datasphere. A July 2026 SAP community series building a knowledge-graph agent on BDC data products puts it in one line: vector databases help AI find things that are similar; knowledge graphs help AI understand how things are connected. Take the question "if work center Z_ASM3 goes down, which customers are affected and what revenue is exposed?" — no similarity search follows the chain work center → product → sales order → customer, and no graph should sum revenue.

Why it matters

• Relationship questions (which customers are hit if a work center fails) cannot be answered by similarity search; aggregations should not be done in a graph.
• HANA Cloud runs vector search, SPARQL and SQL in one engine, so results are merged in the database rather than by the LLM.
• Approximate HNSW indexes, embedding model choice and cross-encoder re-ranking directly change which evidence the agent sees.

Key points

• Vectors find similar; graphs know connected — combine GraphRAG and vector RAG (SAP Learning).
• REAL_VECTOR / HALF_VECTOR, dimension 1–65,000; no row tables, no ordering, not a partition key.
• Similarity: COSINE_SIMILARITY and L2DISTANCE; in-database embeddings with VECTOR_EMBEDDING (NLP service).
• Documented models: SAP_NEB.20240715 and SAP_GXY.20250407 (768 dimensions).
• HNSW index only: M 64, efConstruction 128, efSearch 256 by default; build ONLINE; NO_VECTOR_INDEX for exact search.
• CROSS_ENCODE with SAP_CER.20250701 re-ranks top candidates.
• Knowledge graph engine: SPARQL_EXECUTE, SPARQL_TABLE, SQL_TABLE; federation with SQL and vector engines.
• Pattern: master data in the graph, transactions in SQL over BDC data products (virtual tables), text via vectors.

Terms used on this page

HNSW

Hierarchical navigable small world: the approximate nearest-neighbour index type supported by HANA Cloud vector columns.

efSearch

HNSW search parameter: candidates collected per top-k query; higher raises recall and query time.

Cross-encoder

Model that scores query and passage together; used by CROSS_ENCODE to re-rank vector search results.

GraphRAG

Retrieval that follows explicit relationships in a knowledge graph to assemble context for an LLM.

Competency question

Business question a knowledge graph must answer; drives ontology design and serves as a test case.

SHACL

W3C language for validating RDF graphs against shapes (mandatory properties, types).

Virtual table (BDC data product)

HANA Cloud access path to a BDC data product shared to the instance, without a customer-built ETL pipeline.

Sources

1. SAP Learning — Exploring SAP Knowledge Graph (452,000 ABAP tables, 80,000 CDS views, 7.3 million fields)
2. SAP Community — Knowledge Graph Agent on SAP HANA Cloud Series, Part 1: Why Knowledge Graph for SAP Data (Jul 2026)
3. SAP Help — Vector Engine Guide: REAL_VECTOR and HALF_VECTOR Data Types (QRC 3/2026)
4. SAP Help — Vector Engine Guide: Using the VECTOR_EMBEDDING Function
5. SAP Help — Vector Engine Guide: Creating Text Embeddings with NLP (SAP_NEB, SAP_GXY models)
6. SAP Help — Vector Engine Guide: Embedding Text Values in Tables (generated columns, triggers)
7. SAP Help — Vector Engine Guide: Creating Indexes on REAL_VECTOR and HALF_VECTOR Columns
8. SAP Help — Vector Engine Guide: CREATE VECTOR INDEX Statement (HNSW, M, efConstruction, efSearch)
9. SAP Help — Vector Engine Guide: Improving Search Results with a Cross-Encoder (CROSS_ENCODE)
10. SAP Help — HANA Cloud Knowledge Graph Engine Guide: Connectivity Interfaces (SPARQL_EXECUTE, SPARQL_TABLE, SQL_TABLE), QRC 3/2026
11. SAP Help — SPARQL Reference Guide: SPARQL SELECT Queries Using SPARQL_TABLE
12. SAP News — Business AI Innovation Unveiled at SAP TechEd (Nov 2025): HANA Cloud knowledge graph engine auto-generation from metadata (Q1 2026)
13. SAP Community — Converging Every AI Workload on a Single Database (SAP HANA Cloud, Sep 2026)