AI Glossary

A model-driven system that can choose actions, call tools, and iterate toward a goal.

A constrained sequence where one or more agents plan, act, verify, and hand work between steps.

Splitting source material into smaller passages so it can be embedded, retrieved, or placed into context.

The maximum number of input and generated tokens a model can handle in one request.

A numeric representation of content used to measure semantic similarity.

A representative collection of inputs and expected behavior used to measure an AI system.

Connecting a model response to supplied evidence, tools, or authoritative source material.

A validation or policy layer that checks model inputs, outputs, or actions before they proceed.

A fluent model response that contains unsupported or invented claims.

Running a trained model to generate a prediction, classification, or response.

Stored attention keys and values that avoid recomputing earlier tokens during generation.

The elapsed time between sending a request and receiving useful output.

Using a language model to score or compare outputs against a rubric.

A kind of input or output such as text, image, audio, video, or files.

Instructions in untrusted content that attempt to override the intended behavior of an AI system.

Representing model weights with lower precision to reduce memory and often improve inference speed.

Retrieval-augmented generation: finding relevant sources and placing them into the model context before generation.

A model that reorders retrieved results to put the most relevant evidence first.

A response constrained to a known schema, such as a validated JSON object.

A sampling control that changes how concentrated or varied token selection is during generation.

A unit of text processed by a model, often a word fragment, symbol, or punctuation mark.

A model producing a structured request for an external function, API, or application action.

A store optimized for indexing embeddings and finding semantically similar items.