Reference
Data products and enterprise AI glossary
A shared vocabulary for product leaders, data teams, architects, and AI engineers.
Core terms
Definitions are written for product, data, architecture, and AI teams working together.
- Addressable data product
- A data product that consumers can access through a stable, documented interface such as a table, API, stream, feature view, or application.
- Agentic AI
- AI software that can select actions, call tools, maintain state, and pursue a goal within defined permissions and stopping rules.
- Data contract
- A versioned agreement covering schema, semantics, quality expectations, service levels, access, ownership, and change policy.
- Data mesh
- A socio-technical operating model that assigns data ownership to domains while using shared platform capabilities and federated governance.
- Data product
- A managed data asset or service with a defined consumer, owner, interface, quality promise, support model, and measurable value.
- Data product owner
- The accountable person who prioritizes consumer outcomes, service quality, roadmap, adoption, and lifecycle decisions for a data product.
- Data quality SLO
- A measurable target for properties such as freshness, completeness, validity, accuracy, consistency, or availability.
- Discoverability
- The ability for a consumer to find a data product, understand its purpose, and determine whether it fits the intended use.
- Embedding
- A numeric representation that places semantically related content near one another in a vector space.
- Evaluation set
- A stable collection of representative cases and expected outcomes used to compare models, prompts, retrieval, tools, or system versions.
- Federated governance
- Governance in which shared policies and standards are enforced while domain teams retain responsibility for local data products.
- Feature store
- A system that manages reusable machine-learning features for training and online inference with consistency and lineage.
- GraphRAG
- A retrieval approach that uses entities and relationships in a graph to support structured or multi-hop evidence gathering for generation.
- Grounding
- Constraining a model response with retrieved evidence, tool output, or verified context so claims can be traced to sources.
- Human in the loop
- A control pattern in which a person reviews, approves, corrects, or takes over selected decisions or actions.
- Idempotency
- The property that repeating the same operation produces no additional unintended effect, which is essential for retries and event processing.
- Interoperability
- The ability of data products and systems to exchange and interpret data consistently through shared semantics and interfaces.
- Knowledge graph
- A structured representation of entities, attributes, and relationships that supports traversal, rules, and contextual reasoning.
- Lineage
- The trace of where data came from, how it changed, and which downstream products, decisions, or models depend on it.
- LLM guardrail
- A control that constrains inputs, outputs, tool access, policy compliance, or workflow behavior around a language model.
- Model routing
- Selecting a model or deterministic path based on task difficulty, risk, latency, cost, and required quality.
- Observability
- The ability to understand system behavior through traces, metrics, logs, evaluations, and links to business outcomes.
- Output port
- A supported interface through which a data product serves consumers, such as SQL, API, event stream, dashboard, file, or feature endpoint.
- Prompt injection
- Untrusted instructions embedded in user input or retrieved content that attempt to change system behavior or gain unauthorized access.
- RAG
- Retrieval-augmented generation: retrieving relevant evidence and supplying it to a model when producing an answer or action.
- Semantic layer
- A shared business representation of metrics, entities, dimensions, definitions, and relationships above physical data sources.
- Self-describing data
- Data whose schema, semantics, ownership, quality, usage, and change history are documented and machine-readable.
- Vector database
- A system optimized to store embeddings and retrieve items by similarity, often combined with metadata filters and hybrid search.
- Versioned interface
- A published contract whose changes are tracked so consumers can migrate without unplanned breakage.
- Workflow orchestrator
- The component that coordinates state, sequencing, retries, approvals, scheduling, and execution across steps or agents.
Use the glossary as a navigation layer
Move from a definition to the relevant implementation guides.