An analytics platform lead has had MetricFlow in production for a year. Revenue is defined once in YAML, every dashboard agrees, and the dbt MCP server answers metric questions inside Claude. Then the AI platform team asks a different question. What is a customer, which records count, and who agreed that?
That question is where a dbt Semantic Layer vs TopQuadrant comparison gets decided. dbt's layer defines metrics, dimensions, and entities over dbt models and generates the SQL. TopQuadrant's defines the concept first, with a named owner and a validation rule, and maps it to tables second. A metrics layer defines a concept as a join key on a table. It cannot define a concept that has no table.
The dbt Semantic Layer is a well-documented metrics layer with a deterministic SQL engine and an honest benchmark. In its documentation a concept is a join key, an owner is a meta tag or a group, and approval is a pull request. TopQuadrant holds the concept, its owner, its rule, and its version on one governed asset and serves them with every answer.
Metrics, dimensions, and entities, in YAML, over the models in a dbt project. dbt's MetricFlow documentation, read on 22 November 2026, says the engine "operates through YAML files, where a semantic graph links language to data" (About MetricFlow).
dbt's current pages say dbt platform for the hosted product and Catalog for the browsing surface. MetricFlow is Apache 2.0; the service layer and APIs are proprietary: "The dbt Semantic Layer is proprietary to dbt, however some components of it are open-source" (Semantic Layer FAQs).
From dbt's pages, read on 22 November 2026.
dbt also supports an interchange format: "Starting in dbt v1.12, you can also define semantic models using Ossie documents as an alternative to dbt's native YAML configuration." Apache Ossie's ontology specification lets a concept carry an optional iri pointing to "a class or property in an existing RDF or OWL vocabulary" (github.com/apache/ossie). The specification is a draft whose schema may change, and dbt's own Ossie page covers datasets and metrics, where "Each dataset source must resolve to a dbt model" (Ossie semantic models). Neither specification has an owner or an approval field.
"Semantic Layer vs. Text-to-SQL: 2026 Benchmark Update" by Jason Ganz and Benoit Perigaud appeared on the dbt Developer Blog on 7 April 2026 (dbt Developer Blog). It ran 11 questions, each 20 times, against two models, split by whether MetricFlow could reach the answer.
| Question set | dbt Semantic Layer | Text-to-SQL | Source |
|---|---|---|---|
| All 11 questions | 72.7%, both models | 64.5%, both models | dbt Developer Blog, 7 April 2026 |
| In scope of the modelled layer | 100.0%, both models | 62.5% with Sonnet 4.6; 51.2% with GPT-5.3 Codex | Same page |
| Out of scope | 0.0%, both models. dbt then adds: "With just 3 additional models, the Semantic Layer can now answer every question in the benchmark." | 70.0% with Sonnet 4.6; 100.0% with GPT-5.3 Codex | Same page |
On the out-of-scope row: "the Semantic Layer tells you it can't answer. It never returns invalid data. Text-to-SQL will cheerfully give you a wrong number."
The benchmark scores SQL against a fixed answer key. It does not ask whether the definition the layer answered with was the one the business agreed, or who agreed it. Its scope is "what's been modeled", and the fix for a gap is "additional modeling". Our post Semantic Layer vs Text-to-SQL: What a Refusal Is Worth spends its length on that gap.
The decision that matters. Every row is dbt's documentation, read on 22 November 2026 unless dated.
| Question | What dbt's documentation says | Source |
|---|---|---|
| What is a concept? | An entity: "real-world concepts in a business" that "serve as a join key across semantic models". | Entities |
| Can a concept exist with no table? | As a name only. A primary entity "doesn't necessarily have to map to a column in that table and assigning the name doesn't affect query generation." | Semantic models |
| Who owns a definition? | A group: "every group has an owner". Elsewhere owner is a meta key, and "Changes to meta, including at the column level, don't trigger state:modified." | Groups; meta |
| Who approves a definition? | "Once you've committed your changes and another team member has signed off on the pull request, create a deploy job and run it". | How to build a semantic layer, 18 November 2024 |
| What does an agent receive? | list_metrics returns a "Complete list of metric names, types, labels, and descriptions". | Introducing the dbt MCP Server |
| What did it mean on a past date? | Git history, and model versions: "Governance features are model-specific. They don't apply to other resource types". | Model governance |
Read together: dbt has concepts, owners, approval, and versions. Each is real, and each lives somewhere other than the definition. The concept is a key on a table. The owner is a tag, and dbt "treats meta (and tags) as metadata only". Approval is a merged pull request. The version is a commit. dbt's ownership essay, dated 15 July 2025, says the tools need to move "beyond YAML file editing toward more intuitive user experiences" (Who should own the semantic layer?).
dbt's own how-to, dated 18 November 2024, calls the result "this metrics layer". Its benchmark calls it "a structured ontology". Both are fair: the layer models the concepts that have tables. Our post Metrics Layer vs Semantic Layer: The Difference Is a Noun takes the two layers apart. Our post What Is a Semantic Layer? says what the layer has to hold for an agent.
Our position, with the same weight as dbt's evidence above and open to the same test against our live pages, read on 22 November 2026. TQ Data Foundation is the governed semantic layer the metrics file, the notebook, and the agent read from. The concept is modelled first, on open standards, with a named owner and a validation rule. The mapping to tables comes second, as data you can query. Our site calls the result authoritative context: the concepts, relationships, rules, and processes your experts have modelled, approved, and governed in one place.
| Same question | What TopQuadrant's live pages say | Source |
|---|---|---|
| What is a concept? | "Every metric and entity carries its business definition, its owner, its calculation, and its mapping to physical columns." / "Ontology-bounded joins. Only modelled relationships are traversable." | Agent-ready semantic layer |
| Can a concept exist with no table? | The concept lives in the ontology; the mapping is a separate asset: "Explicit mappings. Concept to table to column, held as data you can query." FIBO concepts arrive before any mapping: "Import the Financial Industry Business Ontology and extend it with your own model." | Semantic layer; Agent-ready semantic layer |
| Who owns a definition? | "Owner per definition. Disagreement resolves to a named person and a governed change process." | Agent-ready semantic layer |
| Who approves a definition? | "AI proposes. Your experts confirm. Inference gets you started, and agreement is what makes it true." | Authoritative context |
| What does an agent receive? | "Definitions before generation. The agent resolves what a term means before composing a query." On the way back: "Provenance on the answer. A returned value carries what it means and where it came from." | Agent-ready semantic layer; Semantic layer |
| What did it mean on a past date? | "Version history. See what the metric meant last quarter, and what changed since." / "Reproducible. Re-run a historical question against the definitions in force that day." | Semantic layer; Agent-ready semantic layer |
Where we do not compete: generating SQL from a metric definition, the JDBC, GraphQL, and Python interfaces, and CI validation on every pull request. Those are dbt's, and the benchmark says they work. We publish no benchmark of our own. What we hold is what the benchmark does not test: the definition the layer answered with, who agreed it, and when. A dbt project keeps its calculations. The concept, its owner, and its rule sit above the tables those calculations run on.
Start with the question that triggered it.
Above the metric. We would be the wrong answer to the first problem. If three dashboards disagree on revenue, dbt's Semantic Layer settles it, and its benchmark shows the SQL is right inside its scope. If an agent is asking what a customer is, which records count, and who agreed, the thing to add is a governed record of the concept above the tables. Our agent-ready semantic layer page shows what that record holds.
Keep the metrics in MetricFlow. Put the concept, its owner, and its rule on a governed asset with a name on it. dbt tells you when it cannot answer. TopQuadrant tells you what the answer rests on.