Structured Outputs enforce a specific response shape, often with schemas, so AI output can be parsed and consumed reliably by software.

Reasoning Effort is a controllable depth setting for model thinking, balancing answer quality, latency, and cost.

A Context Window is the maximum amount of tokens a model can process at once, including instructions, conversation history, and retrieved data.

Tool Calling is the capability for a model to decide when to call connected tools, then use tool results to complete a task.

Retrieval-Augmented Generation (RAG) enriches model outputs by fetching external knowledge at runtime and conditioning generation on it.

Hallucination refers to model outputs that sound plausible but are unsupported or false, often due to missing or weak source context.

Guardrails are technical and procedural constraints, such as filters and validators, that enforce safe and reliable AI behavior.

Related AI terms: Model Card and Content Credentials.

Evaluations (Evals) are benchmarked test suites used to track model behavior, regressions, and improvements against defined criteria.

Human-in-the-loop is an operating model where people validate, correct, or approve AI outputs before final execution.

Model Routing is the strategy of directing requests to the most suitable model dynamically using policy and workload rules.

Related AI terms: Partner Models and Model Card.

Partner Models are external foundation models exposed through a single product interface. Product teams often combine them with Model Routing and governance artifacts like a Model Card.

An MCP host is the application that runs one or more MCP clients, manages connections, and makes responses from connected MCP servers available to the user or model.