Agent. A system in which model output helps select subsequent steps toward a goal, within application-defined controls.
API. An interface through which software requests another system's data or operations.
Authentication and authorization. Establishing who an actor is, then determining what that actor may access or do. Approval expresses intent for a particular action.
Connector. A standing link that lets an AI application read or act on another system, such as email, files or a calendar. A connector usually grants access to far more than the item you had in mind.
Context window. The bounded token capacity available for a model request, subject to the model and service's input and output rules.
Data residency. Where a provider stores and processes data. It affects confidentiality obligations, regulatory exposure and who may compel access.
Embedding. A numerical representation used to compare or process items. Similarity between embeddings does not establish truth or permission.
Evaluation. Repeatable assessment against defined tasks, expected behaviours and failure criteria.
Fine-tuning. Further training of an existing model, changing weights or additional trainable parameters for a purpose.
Function calling (tool calling). A mechanism by which a model proposes a named operation and its arguments in a structured form. The application decides whether to execute it.
Grounding and hallucination. Grounding links claims to supplied evidence. Hallucination refers to generated claims that are false or unsupported.
Idempotency. A property used to ensure that repeating the same intended operation does not create additional effects.
Inference. Running a trained model on input to produce output. Ordinary inference does not update model weights.
KV cache. Stored attention keys and values reused during transformer generation to avoid recomputing earlier state.
LLM. A large language model trained to process and generate language. Many current systems also handle other modalities.
MCP. Model Context Protocol, a standard interface for exchanging context and capabilities between compatible clients and servers.
Memory (product feature). A product capability that stores information from earlier interactions and reinserts it into later requests. It lives outside the model's weights and can carry stale or wrong facts forward.
Model gateway. An application-side service that centralizes provider selection, routing, usage limits and approved model versions. It does not by itself enforce record-level authorization.
Open-weight model. A model whose trained parameters are published for download. Open weights do not imply unrestricted licensing or that training data and code are available.
Parameters and quantization. Parameters are learned numerical values. Quantization lowers numerical precision to reduce storage or computation, with workload-dependent effects.
Prompt injection. An attempt to redirect model behaviour through input. Indirect injection comes through material such as documents or tool results.
RAG. Retrieval-augmented generation, which supplies selected evidence to a model at request time.
Reasoning mode. A setting that gives a system more inference effort on difficult tasks. Reasoning shown to the user explains the answer. It is not a forensic record of internal computation.
Shadow AI. Use of AI services or accounts that the organization has not approved, typically because approved tools are absent, slower or more restricted.
Structured output. A response constrained to a defined shape, such as JSON fields with allowed values. Valid shape does not establish correct content.
System of record. The authoritative source for a business fact, such as an HR system for leave balances. Generated text describes it. It does not replace it.
Token and tool. A token is a unit processed by a model. A tool is a capability the application can execute, often following a model's proposed call.