Working Vocabulary Defines 6 Data Terms for Deterministic and Probabilistic AI Outputs
Updated
Updated · O'Reilly Media · Sep 17
Working Vocabulary Defines 6 Data Terms for Deterministic and Probabilistic AI Outputs
3 articles · Updated · O'Reilly Media · Sep 17
Summary
Six core terms—semantic layer, ontology, knowledge graph, context, observability and output type—are recast into a practical vocabulary aimed at teams deploying AI and data systems.
The framework starts by separating deterministic outputs, which should return the same traceable answer every time, from probabilistic outputs, where LLM responses can vary across runs.
A semantic layer is defined as software for governed metrics and repeatable answers, while an ontology is an entity-first model and a knowledge graph is that model populated with real nodes and relationships.
Context is split between governed metadata and runtime prompt material, and observability is framed as telemetry for catching quiet failures such as drift, retrieval errors and changing model behavior.
The broader argument is that these concepts are layered rather than competing products, helping buyers and practitioners cut through vendor-driven terminology drift.
If probabilistic AI requires rigid semantic layers to be trusted, does this strict control destroy the very reasoning capabilities we initially wanted?
Are enterprises over-engineering an impossible architecture, or is this complex layered approach the only way to stop AI from hallucinating critical business metrics?
When an AI confidently reports the wrong revenue, which layer of this complex new data stack is actually to blame?