What Is a Semantic Layer?

One definition, every consumer

A semantic layer is the layer of software between your warehouse and everything that reads from it. It does three jobs:

1. Names the business vocabulary — entities (Customer, Order), measures (revenue, order_count), dimensions (country, product_line, week).
2. Maps that vocabulary to physical SQL — the revenue metric is SUM(orders.amount_cents)/100, joined to customers on customer_id.
3. Compiles consumer queries against the vocabulary into warehouse-native SQL on demand — Snowflake, BigQuery, Databricks, Postgres.

The leverage: every consumer (Looker, Hex, Streamlit, an embedded app, an LLM agent, a Slack bot) reads the same definitions. 'Revenue' means one thing across the company.

Without a semantic layer

Each tool re-writes its own SQL. Marketing's dashboard says $2.3M; Finance's says $2.1M; the CEO loses faith in both. Three months are spent reconciling the difference, only to find one excluded refunds and the other excluded VAT.

Where it sits

Sources → Warehouse (Snowflake/BigQuery/Databricks)
       → dbt models (raw → staging → marts)
       → SEMANTIC LAYER  ← *this track*
       → Consumers (BI, notebooks, apps, LLMs)

It is not a warehouse, not a dashboard tool, not a metrics dashboard product — it is the typed contract between warehouse and consumer.

A semantic layer is a trusted hotel concierge. Guests (BI tools, notebooks, LLMs) ask in plain language — 'a quiet room near the lobby', 'how was revenue last quarter?' — and the concierge translates that intent into the hotel's internal codes (room numbers, SQL joins) and returns the answer. Two guests who ask the same question always get the same answer because there is only one concierge desk. Without one, every guest wanders the corridors guessing which door is theirs — and they all come back with different rooms.

Where the semantic layer sits — between warehouse marts and every consumer surface. One definition, many frontends.