Reference

AI.

A short reminder. Follow a lesson for the decision behind the term.

LLM

A language model that generates output from learned patterns and supplied context.

Why a PM cares: Fluent output needs evidence and evaluation before it becomes a reliable product promise.

A unit of content processed or generated by a model.

Why a PM cares: Token volume affects context limits, processing cost, and often latency.

The model-specific limit on the material available during generation.

Why a PM cares: Long conversations and documents need explicit handling for truncation and missing context.

Model output constrained to a defined data shape.

Why a PM cares: It simplifies integration but does not guarantee the values or decisions are correct.

A model requesting a software operation and receiving its result.

Why a PM cares: The application must enforce permission, validate targets, and distinguish requests from confirmed outcomes.

Retrieving relevant information and supplying it as context for generation.

Why a PM cares: Evaluate retrieval and answer grounding separately; permissions and source quality remain essential.

A numerical representation used to compare content, often by semantic similarity.

Why a PM cares: Similar meaning can improve search but does not enforce exact identifiers, freshness, or access rules.

A process where a model chooses subsequent actions based on observations toward a goal.

Why a PM cares: Autonomy adds flexibility and failure modes, so scope, budgets, and stopping rules matter.

A repeatable assessment of an AI system against defined tasks and quality criteria.

Why a PM cares: It makes regressions and important failure categories visible beyond a few good demos.

A person assigned a defined review, approval, or exception-handling role.

Why a PM cares: Review helps only when the person can detect the relevant error and act before harm occurs.