Support for Ontologies

[jochenchrist]

Idea

Today the OSI spec lets us describe semantic models — datasets, fields, relationships, and metrics tied to physical tables and columns. This works well for BI and analytics interoperability, but it means the business meaning is always coupled to a specific physical layout.

I'd like to propose adding an ontology layer that describes business concepts (Customer, Order, Product, etc.) independently of where the data lives. Think of it as the "what things mean" layer sitting above the "where things are stored" layer.

Why?

• Multiple physical representations, one meaning. A "Customer" concept might be spread across a CRM table, a data warehouse dimension, and an event stream. The ontology defines it once; semantic models map it to physical sources.
• Reusable across tools. An ontology defined in OSI could feed into knowledge graphs, data catalogs, governance tools, and documentation, not just BI.

What it could look like

At a high level, an ontology block as a peer to semantic_model, containing (and adopting some of Palantir's Ontology ideas):

• Concepts — named business entities with typed properties, primary keys, and inheritance
• Shared properties — reusable property definitions (e.g., createdAt) that concepts reference via ref
• Groups — lightweight organization of concepts into categories
• Relationships — semantic connections between concepts (hasMany, belongsTo, references) distinct from FK-based dataset relationships

A document could contain a semantic model, an ontology, or both. Properties could carry classification labels (e.g., PII) and ai_context hints, following existing OSI conventions.

Example

version: "0.1.1"

ontology:
  - name: commerce
    description: Commerce domain ontology
    concepts:
      - name: Customer
        description: A person or organization that purchases goods
        ai_context:
          synonyms: ["client", "buyer", "account"]
        primary_key: [customerId]
        properties:
          - name: customerId
            logical_type: string
            required: true
          - name: email
            logical_type: string
            required: true
            classification: PII
          - name: isActive
            logical_type: boolean
      - name: Order
        description: A purchase transaction placed by a customer
        primary_key: [orderId]
        properties:
          - name: orderId
            logical_type: string
            required: true
          - name: status
            logical_type: string
            examples: ["pending", "shipped", "delivered"]
          - name: totalAmount
            logical_type: number


    relationships:
      - name: customer_orders
        from: Customer
        to: Order
        type: hasMany
        description: A customer can place multiple orders

Questions for the community

1. Is separating "business meaning" from "physical mapping" a useful distinction for your use cases?
2. Are there other ontology features you'd want to see (e.g., concept-to-dataset mapping, constraints, enums)?
3. Any concerns about scope or overlap with the existing spec?

Would love to hear your thoughts!

Then, I can draft a PR to discuss this on an example and Spec level.

[willpugh]

There is a working group on Ontologies, if you go to the slack channel #osi-ontology it is where you can find the folks working on it.

The current working document is OSI Proposal Ontology

You should feel free to jump on in, and give feedback!

[jochenchrist]

OK, thanks, I will join the channel. And good to know we have some common ideas here :)