Semantic Modelling

01 What it is

Two meanings, one goal: shared meaning

The term carries two complementary senses. In the BI world, the semantic layer is a governed abstraction that maps business terms — "Revenue", "Active Customer" — onto physical tables and columns, so people query concepts rather than hand-writing joins. In the knowledge world, an ontology or knowledge graph is a formal model of entities and their relationships (often in RDF/OWL) that supports inference and linking.

Both exist for the same reason: one consistent definition of everything. That means dashboards that agree with each other, and a grounding layer that keeps LLMs honest. dbt's own benchmarks show natural-language queries answered through a semantic layer reaching roughly 83% accuracy versus ~40% for raw text-to-SQL — the semantic layer is fast becoming core AI infrastructure, not just a BI convenience.

02 Core concepts

The vocabulary of meaning

• Semantic layer — a business-meaning map over physical data.
• Metrics / headless BI layer — metric definitions decoupled from any one BI tool and served to all consumers via API.
• Dimensions vs measures — dimensions group and filter (date, region); measures aggregate (revenue, count).
• Ontology vs taxonomy vs glossary — formal classes and axioms; a hierarchy; an agreed set of terms — increasing formality.
• Knowledge graph — entities as nodes, relationships as edges, plus attributes — queryable and inferable.
• RDF, RDFS, OWL — triples, schema, and rich logic/inference for the semantic web.
• SPARQL & SHACL — the query language for RDF, and a language to validate graph shapes.
• Business-glossary linkage — every metric traced back to an authoritative term definition.
• Embeddings vs symbolic — statistical vectors vs graph/logic; GraphRAG combines both.

03 Key personas

Who builds semantics

04 Methodology

Model meaning, don't guess it

Whether you're building an ontology or a metrics layer, start from competency questions: what must this model be able to answer? Formal methods like METHONTOLOGY and NeOn (reuse-driven networks of ontologies) codify the ontology route; the BI route maps a dimensional model to governed metrics.

05 Best practices

What good looks like

06 2026 trends

Semantics becomes AI infrastructure

↻ Refreshed 2026-07-16 by the AI desk

07 In practice

Where meaning pays off

A semantic layer is quietly becoming core AI infrastructure. Here's how enterprises use it, what the analysts and platforms are doing, and how SCIKIQ delivers it.

What the leaders are doing

The whole industry is converging on a shared semantic layer — capped by the Open Semantic Interchange (OSI), a vendor-neutral spec launched in September 2025.

08 Developer lab

Hands-on: define once, serve everywhere

# dbt / MetricFlow: define a metric once
metrics:
  - name: weekly_active_customers
    type: simple
    label: "Weekly Active Customers"
    type_params:
      measure: distinct_customers
    filter: "{{ Dimension('order__is_active') }}"

semantic_models:
  - name: orders
    model: ref('fct_orders')
    entities:
      - { name: order_id, type: primary }
      - { name: customer, type: foreign, expr: customer_id }
    dimensions:
      - name: ordered_at
        type: time
        type_params: { time_granularity: day }
      - { name: status, type: categorical }
    measures:
      - { name: order_total, agg: sum, expr: amount }
      - { name: distinct_customers, agg: count_distinct, expr: customer_id }

metrics:
  - name: weekly_active_customers
    type: simple
    label: "Weekly Active Customers"       # one definition, org-wide
    type_params: { measure: distinct_customers }

// Build a tiny knowledge graph...
CREATE (c:Customer {name: 'Acme'})-[:PLACED]->(o:Order {amount: 5200})
       -[:CONTAINS]->(p:Product {line: 'Packaging'});

// ...then answer a multi-hop question the semantic layer can't
MATCH (c:Customer)-[:PLACED]->(o:Order)-[:CONTAINS]->(p:Product)
WHERE p.line = 'Packaging'
RETURN c.name, sum(o.amount) AS revenue_at_risk
ORDER BY revenue_at_risk DESC;

# Ground an LLM answer in a governed metric, not guessy SQL
from dbtsl import SemanticLayerClient

client = SemanticLayerClient(environment_id=env, token=token)

# The LLM maps the question -> metric + grain; the layer writes the SQL
df = client.query(
    metrics=["weekly_active_customers"],
    group_by=["metric_time__week"],
    order_by=["metric_time__week"],
)
print(df.head())   # every tool that asks gets the same number

The metric is defined once in YAML, the graph answers relationship questions SQL struggles with, and the Python client lets an agent ask in business terms — all resolving to the same governed definition.

09 Analyst lab

Hands-on: make every metric mean one thing

10 Career path

From analyst to data architect

Semantic skills sit at a premium because they bridge business meaning and technical modelling. Bands are indicative US figures and vary by market.

11 Skill matrix

Rate yourself, then level up

Find your current row and aim one column right.

12 Glossary

Terms worth knowing

13 Test yourself

Check your understanding

14 Keep learning

Where to go next

Semantics needs a solid physical model underneath and clean data flowing in. Continue the path: