Committed ⭐️⭐️⭐️

Question

Semantic relationship / Ontology capability

4 replies
50 views

Whether CDF natively supports semantic/ontology-based relationship modelling (named, typed relationships like contains, hasTimeSeries, performedOn with direction/cardinality/metadata), and whether Atlas AI and other consumers can traverse that semantic meaning rather than just node-to-node links.

data modeling

kelvin

Expert ⭐️⭐️⭐️⭐️

Native Semantic Relationship Modeling

Yes, Cognite Data Fusion (CDF) natively supports semantic and ontology-based relationship modeling via Data Modeling:

Edges: Fully support named/typed relationships, explicit directions, unbounded cardinality (many-to-many, DAG/tree structures), and custom metadata/properties [Direct relations vs. edges].
Direct Relations: Lightweight foreign-key-style links on nodes. They carry semantic meaning (referencing target nodes) but cannot hold properties of their own.

Consumption & Semantic Traversal

Standard Consumers: Fully support querying, traversing, and filtering by typed edges and their properties.
Atlas AI Agents: Atlas AI, powered by the Query tool, or the sandbox, can traverse direct relations, edges or query edge properties. The Query tool only supports edges with Views.

D

Dariush_IDIAuthor

Committed ⭐️⭐️⭐️

Thanks @kelvin , FYI @Anwar Ali Mohamed and @Chandan Kumar

A

Anwar Ali Mohamed

Seasoned ⭐️⭐️

Hi @kelvin ,

We understand that CDF Data Modeling supports Nodes, Views, Direct Relations and typed Edges, including direction and relationship properties, and that Atlas AI can query and traverse these relationships.

We would like to understand whether CDF provides a native ontology/semantic layer beyond this graph structure.

For example:

Compressor isA Rotating Equipment isA Equipment
Work Order performedOn Equipment
Equipment partOf System
Document describes Equipment

Can CDF formally define and govern semantics such as:

Class hierarchy and inheritance
Domain/range of relationships
Inverse relationships
Relationship hierarchy and transitivity
Semantic equivalence between concepts
Cardinality and semantic constraints

More importantly, can Atlas AI reason over these semantics and derive meaning/inferences, rather than simply traversing named relationships that already exist in the graph?

For example, if Compressor isA RotatingEquipment isA Equipment, or a WorkOrder performedOn Compressor and that Compressor partOf CompressionTrain, can Atlas AI use these semantics to infer broader relationships and answer questions at the Equipment or Compression Train level?

Similarly, can it understand that a Document describes an Equipment and a TimeSeries measuresPropertyOf the same Equipment and use these semantics for cross-domain reasoning?

We are trying to understand whether these are merely typed relationships that Atlas AI can traverse, or semantic relationships whose meaning and inference rules are natively understood by CDF/Atlas AI.

The distinction we are trying to clarify is:

A knowledge graph defines how objects are connected. An ontology additionally defines what those objects and relationships mean and enables consistent semantic reasoning.

If CDF does not currently provide a native ontology layer, is the recommended approach to implement this meaning through data-model design, naming/descriptions, relationship definitions and Atlas AI agent instructions, or is there another Cognite-recommended architecture?

This is important for building a governed industrial knowledge graph that can be consistently understood and reused across multiple Atlas AI agents and CDF applications.

@Chandan Kumar @Dariush_IDI

kelvin

Expert ⭐️⭐️⭐️⭐️

Hi @Anwar Ali Mohamed CDF lets you express much of an ontology in the data model, but it doesn’t automatically derive new relationships from rules (for example, “partOf is transitive”).

What you can express:

Node and edge types: both have a type, which is itself a node (for example flowsTo or performedOn).
Edge properties: edges carry their own properties, with direction and target set on the view.
Class hierarchy: views can implement other views, so querying Equipment returns compressors.
Relationship meaning: for example, CogniteMaintenanceOrder.mainAsset effectively says “performed on.”

There is no built-in semantic vocabulary or rule engine. Multi-hop questions are answered by query traversal (or the asset hierarchy), by Atlas AI planning the hops from the model’s names and descriptions, or by writing derived relationships to the graph yourself.

@Uzair Wali can share more on where we’re taking this.

Native Semantic Relationship Modeling

Consumption & Semantic Traversal

Sign up

Welcome to Cognite Hub