Abstain

Report what could not be established when applicable evidence is missing, instead of guessing. Grounded Answers and Abstention

Agent

A loop where a model plans steps and calls tools, with budgets and stopping rules. What an Agent Is

Approval gate

An explicit human approval of one exact action, enforced by the host or executor outside the model. A prompt or a path-string check is not a sandbox. Guardrails and Human in the Loop

Attention

The mechanism that weighs earlier tokens when predicting the next one, with cost growing on long inputs. Context Window and Attention for Users

Cache

Stored results reused across calls, such as shared prompt prefixes, to save tokens and time. Reducing Tokens

Checkable explanation

Assumptions, supporting sources, formulas and intermediate quantities needed to review a result. This is not access to a model's private internal reasoning. Examples and Reasoning

Chunk

One retrievable passage stored with its document, revision, section and source identity. Preparing Engineering Documents

Citation

A source locator (document, revision, section, chunk) attached to the claim it supports. Grounded Answers and Abstention

Context window

The input plus output token budget for one call, including instructions, content, and reply. Context Window and Attention for Users

Embedding

A numeric vector standing for a text span, used to find similar passages for retrieval. Embeddings and Vector Storage

Evaluation

Scoring a tool or prompt on a fixed labelled set before trusting it on real work. Evaluating AI Tools for Your Team

Evidence

A traceable source observation supporting a particular claim, with scope and version retained. Grounded Answers

Few-shot

A prompt that includes a small set of input and output pairs to fix the format. Examples and Reasoning

Fine-tuning

Further training of a model on a narrow dataset to shift its behaviour on that task.

Function calling

A tool-use style where the model fills a named function schema for the code to execute. Tool Use and Structured Output

Guardrail

A tier, approval, limit, or check that bounds what an agent may do unsupervised. Guardrails and Human in the Loop

Hallucination

Fluent output that is invented or wrong, such as fake APIs or signal names. Hallucination and Failure Modes

Inference

Running a trained model to produce outputs, as distinct from training it. What a Language Model Does

Latency

Delay before and during a reply, shaped by first-token time and decode speed. Latency and Throughput

MCP

Model Context Protocol: a connection standard through which an application can access server capabilities. It is not itself an agent, RAG strategy or permission grant. MCP and Skills

MCP host, client, server

The host is the assistant application; each of its clients connects to one server that offers capabilities. What MCP Is

MCP tool, resource, prompt

A tool is invoked with arguments, a resource is read by URI, and a prompt is a reusable message template. Clients, Servers, and Capabilities

Multi-agent workflow

A coordinated division of work across agents, with scoped assignments, evidence handoffs and review. More agents do not guarantee better results. Multi-Agent Design

Parameter

One learned weight of a model, counted in aggregate to describe model size. What a Language Model Does

Prompt

The full input sent to a model, including instructions, content, and format demands. Anatomy of a Prompt

Prompt injection

Instructions smuggled in files or logs that steer the model against the user intent. Guardrails and Human in the Loop

Quantisation

Storing weights or activations at lower precision to cut memory and serving cost.

RAG

Retrieval-augmented generation: retrieving approved, version-filtered passages into the prompt, citing them, and abstaining when evidence is missing. It does not retrain the model. The Retrieval Workflow

Regression suite

A fixed set of tasks rerun after each change to hold gains and catch drift. Evaluating and Operating Agents

Retrieval

Fetching only the passages a task needs, after revision and approval filtering, rather than pasting whole logs or files. The Retrieval Workflow

Sampling

Drawing the next token from the predicted distribution rather than always taking the top one. What a Language Model Does

Skill

A reusable procedure with instructions, references and checks for a bounded task. It does not grant new tool permissions. MCP and Skills

SQL

A language for querying structured tables; use exact queries for counts rather than inferring counts from retrieved excerpts. Engineering Data

Structured output

A reply constrained to a schema such as JSON, so code can validate it. Tool Use and Structured Output

Synthetic / hand-derived

Authored teaching fixtures and paper-derived expectations in this course; not model, tool or execution results. Course Hub

System prompt

Standing instructions that frame every reply, such as role, constraints, and format. Anatomy of a Prompt

Temperature

A setting that sharpens or flattens the next-token distribution before sampling. What a Language Model Does

Throughput

Tokens or tasks completed per unit time under streaming, batching, and rate limits. Latency and Throughput

Token

The billing and context unit of text, shorter than a word on average. Tokens and Tokenization

Tokenizer

The rule that splits text into tokens before counting, pricing, or sending. Tokens and Tokenization

Tool call

A model-proposed action with named arguments, validated and executed outside the model. Tool Use and Structured Output

Training

Fitting model weights on data, as distinct from running the finished model. What a Language Model Does

Workflow

A fixed sequence of steps, approvals and branches chosen in advance, as opposed to an agent that chooses its next action. What an Agent Is