Embeddings and Vector Search on HANA Cloud from Analytic Data
As of 2026-09-25
Teaches vector search on SAP HANA Cloud specifically for structured analytic data — master records, transaction summaries, catalog rows — as distinct from document RAG (M330). Grounds every claim in SAP's documented Vector Engine: VECTOR_EMBEDDING's three arguments, the SAP_NEB.20240715 built-in model (768 dimensions, NLP-feature prerequisite), REAL_VECTOR/HALF_VECTOR storage costs (4/2 bytes per dimension, 1-65,000 dimension range), COSINE_SIMILARITY and L2DISTANCE, and the documented pattern of joining vector-search output back to relational business data in one query. Covers exact versus HNSW approximate search as a measured recall trade-off, not a default; shows that JustAsk's own model-selection step (M347) is the identical embed-then-compare mechanism one layer up the stack; and introduces the Knowledge Graph Engine as the complementary tool for relationship, not resemblance, questions (hybrid grounding, concept C110). Three exercises build an embedded sentence, size a vector column, and choose a search strategy with justification. Builds on M333, M346, M347 and M348.
What you will learn
- State VECTOR_EMBEDDING's documented syntax (text, text type, model/version) and the NLP-feature prerequisite for SAP's built-in model
- Calculate a REAL_VECTOR or HALF_VECTOR column's memory footprint from row count and dimension using the documented formula
- Construct a coherent embedded sentence from structured analytic columns, deliberately excluding non-semantic fields
- Choose between exact and approximate (HNSW) vector search based on scale and a measured recall trade-off, not assumption
- Explain how JustAsk's own model-selection step is the same embed-then-compare mechanism taught in this module, and when a Knowledge Graph Engine traversal is needed instead of or alongside vector search
Module overview
Who this is for. M330 covered the HANA Cloud Vector Engine for document RAG. This module covers a different, narrower case a consultant meets constantly once M347 and M348 are in place: turning rows of structured analytic data — master records, transaction summaries, catalog entries — into vectors, and joining the results of a similarity search back to the governed business data that gave them meaning. It builds on M333 (embeddings fundamentals), M346 (AI-ready data) and M349 sits after M348 because it is the retrieval mechanism text-to-SQL and Just Ask cannot replace: finding the similar, not the exact.
Prerequisites
- M333 (AI & LLM Fundamentals) or working knowledge of embeddings and similarity, and M330 (HANA Cloud Vector Engine & Knowledge Graph Engine) if available
- M346 (AI-Ready Data) or understanding of data-quality checks before grounding
- M347 and M348, or working knowledge of JustAsk's semantic-search-then-query pipeline
- SQL proficiency in SAP HANA Cloud, including table design and joins
Outcomes
- Design a vector-enabled table for structured analytic data that joins cleanly back to its governed source.
- Produce a memory-footprint estimate for a proposed vector column before committing to it in a client proposal.
- Choose and justify exact versus approximate search with a measured recall comparison, not a default assumption.
- Explain to a client, precisely, the difference between a similarity search and a knowledge-graph relationship query, and when both are needed together.
Full module available to members. The full module adds: the decision framework · the end-to-end scenario walkthrough · the KPI scorecard · the anti-patterns · the code blocks · the knowledge check · the diagrams.