Concept card

The semantic layer as LLM context — analytic models, business semantics

As of 2026-11-08

What is The semantic layer as LLM context?

A semantic layer was built for humans — it turns "GL account 400100" into "Net Sales" — and it turns out to be exactly what an LLM needs too: governed vocabulary, hierarchies and business rules an agent can reason with instead of guessing from raw column names. Analytic models exposed to Just Ask are Datasphere's version of this, and their quality is now checked automatically before Joule ever gets to use them.

The layer built for people turns out to be the layer AI needs

A semantic layer was never an AI concept. Decades before Joule existed, BI tools built one because a finance analyst should not have to remember that revenue lives in table FAGLFLEXT and that a currency conversion has to run before a number means anything — the semantic layer did that translation once, centrally, so every report used the same vocabulary and the same rules. What has changed is who else now reads that layer: an LLM asked "what were our net sales in Germany last quarter" cannot answer correctly from a raw schema, but it can answer correctly from a model that already defines Net Sales as a measure, Germany as a member of a geography hierarchy, and last quarter as a time dimension with the fiscal calendar baked in. The semantic layer was built for humans and happens to be exactly the retrieval-free, structured context an LLM needs to stop guessing.

Why it matters

A semantic layer is not an AI invention — it is decades-old BI discipline that turns out to be exactly the structured, governed context an LLM needs; treating it as new Joule work rather than reusing existing modelling discipline duplicates effort.
A hierarchy or label gap that a human report author silently works around becomes, for Joule, a defect that degrades a conversational answer — the tolerance for ambiguity is much lower for an LLM than for a human report reader.
The choice between a semantic-layer-mediated path and a text-to-SQL path is a governance trade-off, not a technology preference — knowing which one a use case needs changes the whole delivery approach.

Key points

The semantic layer predates AI: measures, hierarchies, currency logic and business labels built on Datasphere views for human reporting.
What changed is the reader: an LLM needs the same governed vocabulary to avoid guessing meaning from raw column names.
A model exposed to Just Ask becomes reachable by Joule's Analytical Insights through the same path — the semantic layer is what makes that reachability meaningful.
SAC's Q3 2026 model readiness check flags missing semantics, unclear hierarchies and ambiguous labels before they degrade a Joule answer — a linter for meaning, not a data-quality tool.
Semantic-layer-mediated NLQ trades flexibility for governance guarantees, versus direct text-to-SQL grounding on raw schemas.
Semantic-layer design should be reviewed as a Joule-readiness deliverable, re-checked after every structural change to an exposed model.

Terms used on this page

Semantic layer: The governed vocabulary — measures, hierarchies, currency logic, business labels — that sits between raw tables and a report or an AI agent.
Analytic model: The Datasphere/SAC object that carries the semantic layer for a given business area, exposable to Just Ask.
Model readiness check: SAC's Q3 2026 automated pass over a model that flags missing semantics, unclear hierarchies and ambiguous labels.
Text-to-SQL: The alternative approach that grounds an LLM directly on database schemas instead of a semantic layer — see “Natural-language query over SAP data — text-to-SQL patterns and limits”.
Restricted / calculated measure: An analytic-model construct that filters or derives a business number (e.g. Net Sales as Revenue restricted to actuals) so every consumer, human or agent, reuses the same definition.
Hierarchy (analytic model): A parent-child structure (product, geography, organization, time) that lets an agent resolve a member like 'Germany' to the correct level and aggregate correctly above or below it.
Grounding — two SAP meanings: In the orchestration service, grounding is a RAG pipeline module retrieving documents; in Joule/analytic-model terms, grounding means resolving a question against a governed semantic layer or the Knowledge Graph — the two are not interchangeable.
Cortex Analyst semantic model (peer term): Snowflake's YAML-declared equivalent of a Datasphere analytic model: logical tables, dimensions, measures and synonyms that Cortex Analyst must consult before answering.

Sources

Full card with the Consultant plan — €19.90/month. What the full card adds: the full decision framework · the SAP vs Snowflake / Databricks / Fabric comparison · the common pitfalls and their fix · the cheat sheet · the architecture schemas · the code blocks · the facts worth quoting.

Open in the app →

Edited by Cédric Mary · How this page is made

Open missions

102 live roles and contracts in SAP AI & Joule right now — see the SAP AI & Joule missions · freelance SAP contracts.