Microsoft Fabric Ontology Semantic Layer: Architecture, NL2Ontology & Palantir Contrast

Microsoft definition of Ontology

Ontology is the digital representation of an enterprise vocabulary and semantic layer that unifies the meaning across domains and OneLake data sources.

Key elements: (1) enterprise vocabulary standardizes business terms such as customer, order, and product across the organization; (2) unified meaning for data stored in OneLake, which may include lake‑warehouse tables, event streams, and Power BI semantic models.

Core concepts for developers

Entity Type : reusable logical model (e.g., Shipment, Product, Sensor) that locks the definition of a term.

Entity Instance : concrete occurrence of an Entity Type populated through data binding; records provenance.

Property : named fact about an entity with a data type, enforcing consistent types, units, and naming.

Relationship : typed directed link between entities, optionally carrying attributes (distance, confidence, effectiveAt) and cardinality rules (a customer can have many orders).

Data binding maps entity definitions to real OneLake data: keys to fields, columns to properties, and cross‑source relationships. It also defines schema‑evolution rules, data‑quality checks (non‑null, range, uniqueness), and lineage. After binding, raw rows and events become governed business objects.

Bound objects form an Ontology Graph where nodes are entity instances and edges are asserted or derived links with metadata, supporting visual exploration, graph algorithms (path, centrality, community) and rule‑driven reasoning.

NL2Ontology query layer

The system automatically routes queries to the appropriate engine: graph‑related queries to GQL, time‑series queries to KQL. NL2Ontology converts natural‑language questions into structured queries, ensuring that filters, joins, units, and time windows align with ontology definitions. This creates a declarative semantic layer focused on analysis rather than write‑back actions.

Side‑by‑side comparison with Palantir Foundry

Data source : Microsoft uses OneLake (lake‑warehouse, event streams, Power BI); Palantir integrates multiple sources (ERP, CRM, IoT, documents).

Query method : Microsoft employs NL2Ontology + GQL/KQL; Palantir uses object queries and graph traversal.

Agent integration : Microsoft combines Fabric IQ Agent with external agents; Palantir uses OSDK + Ontology MCP.

Action execution : Microsoft’s scope is limited to querying; Palantir supports full transaction write‑back and edge synchronization.

Security model : Microsoft provides tenant‑level permissions; Palantir offers fine‑grained object‑level permissions.

Maturity : Microsoft Fabric is in Preview; Palantir Foundry is generally available.

Concrete example

Question: “What are the top three SKUs sold in Shanghai stores last month, and who are their suppliers?”

Map “Shanghai store” to the store entity and apply a region filter.

Map “last month” to a time window.

Translate “top selling” into a sales aggregation with a Top‑3 limit.

Translate “supplier” into a traversal from SKU to its supplier entity.

Automatically select GQL to run the graph query and return results using business terminology.

The user never sees the underlying tables or query language; the answer respects the company‑wide definitions.

References

https://learn.microsoft.com/en-us/fabric/iq/ontology/overview

https://learn.microsoft.com/en-us/fabric/iq/

https://geneontology.org/