Embeddings and vector search from Datasphere and HANA Cloud tables

Where the vector capability actually lives

A common misreading is to look for "Datasphere's vector store" as a distinct product. It is not one: the vector capability sits in SAP HANA Cloud, the database Datasphere is built on top of (C001). SAP's own description of the HANA Cloud vector engine names three things it does: generate a vector embedding from text, run a nearest-neighbour search using cosine similarity expressed directly in SQL, and enrich the results by joining the vector-search output with ordinary relational business data. That third point is the one worth sitting with: a similarity search here does not hand back an isolated list of matches the way a standalone vector database would — it returns rows that can be joined, in the same query, to the measures, hierarchies and business context that already live in a Datasphere-modelled table.

Why it matters

• SAP's vector engine joins similarity-search results back to relational business data in the same SQL query — a structural difference from a standalone vector database that changes what "grounding on your own tables" can deliver.
• Embedding structured business data (this card) and document grounding (C116) are different pipelines with different tooling; scoping them as the same work under-delivers on both.
• Where an embedding lives — HANA Cloud versus a Databricks-side vector search (C351) — is a governance decision as much as a technical one, since it decides whether a second, ungoverned copy of the data gets created.

Key points

• The vector capability sits in HANA Cloud, the database beneath Datasphere — not in a separate Datasphere-specific vector product.
• SAP names three vector-engine capabilities: text-to-embedding generation, nearest-neighbour search via cosine similarity in SQL, and joining vector results back to relational business data.
• This card covers embedding structured table data (product descriptions, feedback, incident notes); document grounding (PDFs, SharePoint) is a separate pipeline (C116).
• Keeping embeddings inside HANA Cloud preserves governance — the vector stays in the same access-controlled place as the row it was derived from.
• SAP pairs the vector engine with a knowledge-graph engine on the same database for hybrid retrieval (C110): traversal, SQL aggregation and vector similarity without moving data.
• Choose HANA Cloud vs. Databricks Mosaic AI (C351) vector search based on where the consuming application runs, not on raw capability.
• Mixing outputs of two different embedding models in one index silently breaks similarity search (C154) — a common, avoidable design error.

Terms used on this page

HANA Cloud vector engine

SAP HANA Cloud's capability to store real vector data, generate text embeddings, and run cosine-similarity nearest-neighbour search in SQL, joinable to relational data.

Structured-data embedding

Turning a text column already stored in a governed business table into a vector alongside its row, as distinct from embedding documents.

Hybrid grounding

Combining vector similarity search with knowledge-graph traversal and SQL aggregation on the same database (C110).

Embedding model mismatch

The error of mixing vectors from two different embedding models in one index, which silently breaks similarity comparisons.

Sources

1. SAP — Run semantic search on data with the SAP HANA Cloud Vector Engine (text embedding, cosine-similarity nearest-neighbour in SQL, joins to relational data)