An architect at an insurer gets the same request from two teams in one week. The analytics team wants a semantic layer, so every dashboard computes loss ratio the same way. The AI team wants a knowledge graph, so the claims agent can tell a policyholder from a claimant. Each team says the other's thing is already covered. The semantic layer vs knowledge graph question is whether they are asking for one asset or two.
They are asking for two jobs that share one record. A semantic layer is where a business keeps what its data means and serves that meaning to every tool that queries it. A knowledge graph is a structure: entities, typed relationships, and identifiers, in a form a machine can traverse. Meaning lives in the layer's definitions. The graph is where those definitions, and the things they describe, are stored and connected.
Meaning lives in the semantic layer's definitions. Those definitions, and the entities they describe, are stored and connected in a knowledge graph. One governed record can do both jobs when it is held on open standards with a named owner.
One word is about meaning and who agreed it. The other is about structure and how it is held. Our post What Is a Semantic Layer? defines the layer and the eight things it holds. This article takes the one row of that definition the market gets wrong: where the meaning is kept.
| Dimension | Semantic layer | Knowledge graph |
|---|---|---|
| What it is | A job: serving agreed meaning to every consumer | A structure: entities and typed relationships a machine can traverse |
| Unit of record | A definition, of a concept or a metric | An entity, or a relationship between two |
| Question it settles | What does net revenue mean, and how much is it? | Which customer holds which policy under which contract? |
| Where the meaning is written | Whatever the product speaks: often a YAML file or a database object | In the data, on RDF, OWL, and SKOS, validated with SHACL, the W3C Shapes Constraint Language |
| How it fails | Three numbers for one metric | A populated graph nobody approved |
A definition is a thing with an owner, a version, and relationships to other definitions. That is a node. A semantic layer built on an ontology, which is a formal model of what the business's concepts are, is already a graph. Our post Semantic Layer vs Ontology: Which One Says What a Customer Is takes the ontology half.
The metrics layers most people meet first keep their definitions as configuration. dbt's documentation on semantic models, last updated 2026-09-08, says to "Configure semantic models in a YAML file within your dbt project directory". The same page says to "Think of semantic models as nodes connected by entities in a semantic graph", where an entity is any column "that serves as a join key" (dbt documentation). Snowflake's overview of semantic views, read on 2026-11-22, says "You can store semantic business concepts directly in the database in a Semantic View, which is a schema-level object". The same page says "Semantic views are considered metadata" (Snowflake documentation).
Both descriptions are accurate, and both put the meaning where a machine executes it and a person has to go looking for it. In a YAML file, the definition of a customer is a comment beside a join key. In a knowledge graph it is a node with an owner on it. A metrics layer can define a calculation. It cannot define a concept. Our post Metrics Layer vs Semantic Layer: The Difference Is a Noun takes that sentence apart.
It needs one when the definitions relate to each other and a machine has to follow the relationships. For a layer that only serves metrics to dashboards, a model file is enough. The agent case is different, for three reasons.
At what size does a governed table with an API stop doing the job? Our post on how a knowledge graph relates to the job of serving agreed meaning to an AI system answers that. We call the job context management, and the meaning it serves authoritative context.
No. A populated graph can be accurate and still answer no dashboard. Four things have to be added, and each is a governed asset of its own.
| Added | What the graph has | What the layer adds |
|---|---|---|
| Metrics | The concept an order is | Net revenue as a calculation on that concept, with its grain and dimensions |
| Mappings | An identifier for each entity | The approved path from each concept and metric to the physical columns |
| Serving | A graph query endpoint | The same record answered in SQL, over an API, and over MCP, because BI tools speak SQL |
| Entitlements | Whatever the store enforces on nodes | Rules on the concept, evaluated for each request against the asker's role |
The storage under the graph is a separate question and it stays separate here. Our post Knowledge Graph vs. Graph Database settles what a database holds, what identity across systems needs, and what inference and validation add above storage.
Three vendor pages on this query land on the same split and say so.
So the category agrees that the two stack. What all three do next is the problem. Each hands "what is a customer" to the graph or the ontology and leaves the metric in the layer. That puts the concept and the calculation on two assets, in two products, with no named owner on either. Galaxy names SHACL in one sentence. The other two pages do not name it. None of the three says what happens when the metric moves and the concept does not. That is the meeting from paragraph one, held every quarter.
Here is the difference as a walk. An agent is asked for net revenue from active customers last quarter. Our post What Is a Semantic Layer? shows the governed record for net revenue as a table. In a knowledge graph it is a set of nodes, and the agent reaches each by following an edge.
| Hop | Node reached | What it carries | What the agent does with it |
|---|---|---|---|
| 1. Net revenue | A metric | Definition, calculation, an owner in Finance, version 3, review due 2027-03-14 | Quotes the definition and uses the approved formula |
| 2. measures | Order, a concept | A SHACL shape: exactly one currency, a non-negative gross amount | Knows which records count as orders |
| 3. placed by | Customer, a concept | The scope rule for "active", set by Finance | Applies the agreed filter instead of inventing one |
| 4. mapped to | Two columns in the warehouse | Mappings proposed by automation and approved by the owner | Composes the query over approved columns only |
| 5. visible to | An entitlement | Finance and Sales roles; refund detail to Finance only | Returns what its caller may see, and no more |
Every hop is an edge in the same graph the definitions live in. The join from order to customer exists because the model declares it, so an invented join is unavailable. The answer carries the definition version it resolved, so it can be replayed. Automation can propose most of these nodes. A person with accountability approves them, and our post Ground Truth: SME-Approved vs Inferred Metadata covers what that approval adds. Our post A Semantic Layer for AI Agents: What the Agent Needs Before It Asks for Revenue puts the five things the agent needs in order.
TopQuadrant does not sell the warehouse, the dashboard, or the YAML file. TQ Data Foundation is a semantic layer held as a knowledge graph on open standards. Every metric carries its business definition, its owner, its calculation, and its mapping to physical columns. The meaning and the logic live on the same governed asset. Mappings from governed concepts to physical schemas are explicit assets you can query. Source, model, metric, and report are one connected graph, so a value carries where it came from. Agents receive concepts, permitted relationships, and metric logic over API, SPARQL, and MCP. Only modelled relationships are traversable. Entitlements attach to the governed concept and are evaluated per request. AI proposes the mappings and definitions; your experts confirm them. Every definition is validated with SHACL, the W3C standard whose Recommendation lists a TopQuadrant editor. The standards are open, so the record exports in full. Your catalog stays, and it syncs.
So the two teams in the first paragraph are funding one asset. The semantic layer is what it means. The knowledge graph is where that meaning is kept. TopQuadrant holds both on one record, and a person signed it.