Glossary and cheat sheet

A–C D–H I–P R–W

A/B test — A comparison that assigns comparable users or cases to different versions so their outcomes can be evaluated fairly; see A/B testing.

Agent — Software that combines a model with instructions, context, and permitted tools to pursue a bounded task; see Introduction to AI agents.

API — An application programming interface is a defined way for one program to request information or actions from another; see Connecting to existing systems.

Authentication — A process that checks who a user or system is, rather than deciding every action they may perform; see Authentication and permissions.

Authorisation — The decision about which information or actions an identified user may access in a particular situation; see Authentication and permissions.

Bias — A systematic tendency that can produce unfair or unrepresentative outcomes rather than random mistakes; see Bias, fairness, and responsible AI.

Cache — Stored information or a previous result reused to avoid repeated work, subject to freshness and access restrictions; see Cost optimisation and caching.

Channel — The place where people interact with an assistant, such as a website chat or messaging service; see Integrating channels.

Chunk — A smaller passage made from a larger document so relevant information can be retrieved without sending everything; see Preparing documents for knowledge.

Citation — A reference that lets the reader inspect the source supporting an answer’s particular claim; see Writing good documents.

Context — The instructions, messages, facts, and tool results supplied to the model for its current response; see Adding context.

Context window — The model’s capacity for the tokens considered within a request, with accounting details depending on the model; see Context window, tokens, and cost.

Deflection — An eligible issue resolved without human handling, which should not be confused with an abandoned or blocked conversation; see Metrics.

Deployment — Making a tested version available to its intended users with monitoring, ownership, and a recovery plan; see Deployment checklist.

Embedding — A numerical representation used to compare the meaning or characteristics of text and other content; see Embeddings and semantic search.

Escalation — Moving an issue to a person or process with the authority needed to handle it; see Transparency and human escalation.

Evaluation — A planned assessment against defined examples and criteria, rather than judging an agent from one impressive answer; see Testing and evaluating agents.

Fallback — A defined alternative used when the preferred operation cannot complete, without silently weakening safety or permissions; see Errors, retries, and fallbacks.

Function calling — A model’s structured request for software to execute a named operation with specified inputs; see Function calling.

Grounding — Connecting an answer to supplied evidence instead of relying only on patterns learned during training; see What is a RAG.

Guardrail — A behavioural or software boundary intended to keep an agent within its permitted scope and actions; see Adding guardrails.

Hallucination — An unsupported or incorrect model-generated claim that may sound fluent and confident despite lacking evidence; see Measuring adherence and hallucination.

Human in the loop — A person deliberately involved at a defined decision or approval point rather than merely observing afterwards; see Human in the loop.

Idempotency — A property that lets repeated requests for the same operation avoid creating additional unintended effects; see Errors, retries, and fallbacks.

Inference — Running a trained model on new input to obtain an output, rather than changing the model through training; see What is an LLM.

Instruction — A statement telling an assistant how to behave, which still needs software enforcement for consequential permissions; see Writing good prompts and instructions.

Latency — The time between a request and a useful response, including search, tools, and model processing; see Performance and latency.

LLM — A large language model is a trained system that generates language by predicting successive pieces of text; see What is an LLM.

Log — A recorded event that helps operators understand what happened without needing to reproduce every user interaction; see Logs, traces, and monitoring.

MCP — Model Context Protocol is a shared protocol for exposing tools and resources to compatible applications; see Model Context Protocol.

Memory — Information intentionally retained and later supplied to an assistant, not automatic human-like recollection by the model; see Memory.

Metadata — Descriptive information about a document, such as its owner, audience, or effective date; see Finding and preparing knowledge.

Multimodality — The ability to handle more than one form of information, such as text, images, or audio; see Multimodality.

Prompt — The input that asks the model to do something and supplies relevant guidance or material; see Anatomy of a prompt.

Prompt injection — An attempt to make an assistant treat untrusted content as instructions that override its intended task or boundaries; see Prompt injection and jailbreaks.

RAG — Retrieval-augmented generation retrieves relevant evidence and includes it when asking the model to produce an answer; see What is a RAG.

Rate limit — A restriction on how often requests may be made, which applications must respect rather than bypass; see Errors, retries, and fallbacks.

Reranking — Reordering retrieved candidates using another relevance assessment before choosing which passages to use; see Retrieval strategies.

Retrieval — Finding candidate information relevant to a question within the sources the current user may access; see Retrieval strategies.

Retry — Another attempt after a failed or uncertain operation, subject to limits and duplicate-action checks; see Errors, retries, and fallbacks.

Schema — A description of the required shape and types of data, not proof that the values are factually correct; see structured responses.

Semantic search — Finding content by meaning rather than relying only on exact word matches; see Embeddings and semantic search.

Skill — A reusable instruction bundle for a task or method, distinct from an executable tool; see Adding skills.

Temperature — A setting affecting how model outputs are sampled, not a dial that guarantees factual accuracy; see Parameters.

Token — A small unit of text processed by a model, which may be a word, part of a word, or punctuation; see Context window, tokens, and cost.

Tool — A defined software operation an agent may request, such as searching a database or creating a ticket; see Adding tools.

Trace — A connected record of the steps within an operation that helps locate failures or delays; see Logs, traces, and monitoring.

Webhook — A notification sent to another system when a defined event occurs, often to start follow-up work; see Webhooks, events, and automations.

Workflow — An organised sequence of tasks and decisions that defines how work proceeds and where it stops; see Workflows as skills.