The acronym decoder.

A system where a model dynamically chooses steps or tools toward a goal, within the application’s constraints.

Artificial intelligence: an umbrella term for systems performing tasks associated with intelligent behavior. It is broader than chatbots.

Submitting work under a provider’s asynchronous batch mode. Pricing and completion requirements differ from interactive calls.

A defined task set and scoring setup used to compare performance. Results depend on conditions and may not match your workflow.

A saved execution state that can support inspection or resuming a workflow. It is not a guarantee of safe behavior.

A passage or segment prepared for retrieval. Boundaries affect which context is available around a matched passage.

The model’s finite working context for a request. Available input and output capacity depends on the model and product.

Total workflow spend divided by outcomes meeting the declared acceptance criteria, including failed attempts in the spend.

The policies and responsibilities controlling how data is accessed, used, retained and managed in a system.

A numerical vector representing content so relatedness can be compared for tasks such as search.

A repeatable assessment of outputs or system behavior against explicit criteria and representative examples.

A collection of cases used to check behavior. Preserve versions and avoid treating a small set as universal proof.

Supplying a small set of examples in the prompt to demonstrate the desired input/output behavior.

Further training a model on additional examples to alter its behavior. It is not the same as retrieving documents at request time.

Systems that generate outputs such as text or other media rather than only selecting a fixed label.

Connecting a response to supplied or retrieved evidence. The evidence still needs to support the actual claim.

A control intended to constrain inputs, outputs or actions. It reduces specific risks; it is not proof of universal safety.

A plausible-looking output that is false or unsupported. Fluency alone is not evidence of correctness.

A workflow that includes a person at a defined review, approval or correction point.

A prompt-injection attempt carried in external content such as a webpage or document rather than a direct user instruction.

Using a trained model to compute an output from new input, rather than training its parameters.

Tokens processed as input to a model call; billing may distinguish cached input and other categories.

A model that learns patterns in language and produces or scores token sequences. Its output is not automatically verified.

Large language model. A language model trained at substantial scale; “large” is not a standardized quality rating.

Model Context Protocol: a standard for connecting AI applications with tools, data and other capabilities.

The component that maintains a protocol connection to an MCP server on behalf of a host.

The AI application that manages clients and presents capabilities to the user.

A service exposing capabilities such as tools and resources through MCP. A connection is not authorization for every action.

Documentation describing a model’s intended uses, training context and limitations; details vary by publisher.

The learned numerical parameters of a model. Access to them is distinct from the license governing their use.

Able to process or generate more than one kind of data, such as text and images. Supported combinations depend on the model.

A description of model weights being available to obtain. It does not by itself promise unrestricted use or open training data.

Coordinating steps, state, tools and decisions across a workflow or agent system.

Tokens generated by a model call. Their price can differ from input token pricing.

The instructions and input supplied to a model for a task. Relevant context and examples may be part of it.

Reusing eligible repeated context under a provider’s caching rules. Eligibility, retention and price are provider-specific.

An attempt to redirect model behavior using instructions in input, including external material the application retrieves.

Retrieval-augmented generation: retrieve relevant material, then supply it as context when generating an answer.

Finding candidate evidence or records for a task. A retrieval error can give a model the wrong basis for its answer.

Identifying, measuring and managing risks in context throughout a system’s lifecycle rather than relying on a single safety claim.

Explicit grading criteria defining an acceptable output and the severity of failures.

Search using meaning-related representations rather than relying only on exact word matches.

Information retained across workflow steps, such as intermediate results and pending approvals.

An output constrained to a defined structure, such as a schema. Matching a schema does not establish factual accuracy.

An instruction supplied by the application to guide model behavior, distinct from ordinary task data. Exact roles vary by API.

A unit a model processes, often a word fragment or punctuation in text. A token is not always a word.

Converting input into model-compatible token units. Different tokenizers can split the same text differently.

A request from a model-driven application to an external function or capability. The application controls execution and permission.

A record of execution steps and intermediate activity used for debugging or inspection. It may contain sensitive content.

The process of adjusting model parameters using data and an optimization objective. It differs from adding context to a prompt.

A neural-network architecture widely used in modern language models, with attention mechanisms for processing relationships in sequences.

Retrieving items by similarity between numerical representations. Similarity does not prove truth or freshness.

A sequence of steps with a predefined route. It may contain model calls without becoming fully autonomous.

Requesting a task without worked examples in that prompt. It does not mean the model had no training.