SAP HANA Cloud Vector — RAG patterns

HANA Cloud ships a native vector column type (REAL_VECTOR) and a SQL function VECTOR_EMBEDDING(text, model) that calls AI Core embeddings inline. That means you can ingest, embed, and query without leaving SQL — but the Python clients (langchain-hana / generative-ai-hub-sdk) are usually nicer.

Schema pattern

CREATE TABLE DOCS (
  ID NVARCHAR(64) PRIMARY KEY,
  TEXT NCLOB,
  METADATA NCLOB,                  -- JSON
  EMBEDDING REAL_VECTOR(1536)      -- dim = your embedding model's dim
);

-- Cosine HNSW index — fastest, smallest, default choice.
CREATE HNSW VECTOR INDEX IDX_DOCS_EMB ON DOCS(EMBEDDING) SIMILARITY FUNCTION COSINE_SIMILARITY;

Before ingestion: data preparation

Run sap-hana-data-prep before this skill. It produces the ingestion contract and chunked artifacts this skill consumes:

prepared/chunks.jsonl            -- id, text, metadata per chunk
prepared/ingestion-contract.yaml -- confirms target_shape: hana_vector, embedding_model, pii_handling

If prepared/ingestion-contract.yaml is absent or target_shape is not hana_vector, return to sap-hana-data-prep.

Ingestion via SQL only

INSERT INTO DOCS (ID, TEXT, EMBEDDING)
  VALUES (?, ?, VECTOR_EMBEDDING(?, 'AICORE', 'text-embedding-3-large'));

For region-availability of text-embedding-3-large see sap-sovereign-regions and run recipes/00-develop/00-region-preflight/ — China and KSA need to swap to a multilingual SAP-hosted embedding instead.

Ingestion via Python

from langchain_hana import HanaDB
from gen_ai_hub.proxy.langchain.openai import OpenAIEmbeddings

embeddings = OpenAIEmbeddings(deployment_id=os.environ["EMB_DEPLOYMENT_ID"])
store = HanaDB(connection=hana_conn, embedding=embeddings, table_name="DOCS")

store.add_texts(
    texts=["...chunk 1...", "...chunk 2..."],
    metadatas=[{"source": "doc.pdf", "page": 1}, ...],
)

Query

results = store.similarity_search_with_score("how do I deploy to Kyma?", k=4)

Or as SQL:

SELECT TOP 4 ID, TEXT, COSINE_SIMILARITY(EMBEDDING, VECTOR_EMBEDDING(?, 'AICORE', 'text-embedding-3-large')) AS SCORE
  FROM DOCS ORDER BY SCORE DESC;

Pitfalls

• Wrong dimension → INSERT fails. Pick the embedding model first, then create the table.
• No HNSW index → slow at >10k rows. Add the index after the bulk-load, not before.
• Mixing embedding models without re-indexing → silent quality drop. Pick one and stick to it.
• Embedding model mismatch with region → run recipes/00-develop/00-region-preflight/ first and cross-check sap-sovereign-regions; China / KSA / NS2 constraints on text-embedding-3-large apply here too.

Verify

SELECT COUNT(*) FROM DOCS WHERE EMBEDDING IS NOT NULL;
-- expect: > 0
SELECT TOP 1 ID, COSINE_SIMILARITY(EMBEDDING, VECTOR_EMBEDDING('hello', 'AICORE', 'text-embedding-3-large')) FROM DOCS;
-- expect: a score in (0, 1)

Cross-references

• Upstream data prep: skills/sap-hana-data-prep/ — run this first for any new dataset
• Walkthrough: recipes/optional/hana-vector-store/
• Template: references/kg-project/ for the KG+vector hybrid pattern — planned; tracked as TODO in references/README.md.
• HANA docs: https://help.sap.com/docs/hana-cloud-database/sap-hana-cloud-sap-hana-database-vector-engine-guide