QR-04

AI Glossary Quick Card

Fifty essential terms in plain English. Pairs with the AI Primer track.
Last verified 2026-07-09
How to read this card. Terms are grouped by what you are doing, not alphabetical. If you are new: start with Core. If you build: skip to Prompting, Retrieval, and Agents. If you buy: Evaluation, Cost, and Risk are the ones that will show up in your contracts.

Core Concepts

LLMA model trained to predict text. What you are actually using when you say "AI"
TokenA chunk of text (about 3/4 of a word). What the model reads, and what you are billed by
Context windowThe most text the model can consider at once. Its working memory
EmbeddingText turned into a list of numbers, so a computer can measure similarity
InferenceRunning a trained model to get an answer. The thing you pay for
ParametersThe learned weights inside a model. Loosely: its size
MultimodalHandles more than text: images, audio, video
Foundation modelA large, general-purpose base model that others build on

Prompting

PromptEverything you send to the model. Not just your question
System promptThe standing instruction that sets role and rules for the whole session
Few-shotGiving the model examples in the prompt to shape output
Chain-of-thoughtAsking the model to reason step by step before answering
Structured outputA response constrained to a schema, usually JSON
Context engineeringDesigning what goes into the window. Prompting is one part of it
Prompt cachingReusing computed input to cut cost on repeated prefixes
TemperatureA dial for how much the model varies. Low for facts, high for ideas

Retrieval (RAG)

RAGRetrieval-augmented generation: fetch relevant docs, then answer from them
ChunkingSplitting documents into passages small enough to embed
Vector databaseStorage that finds text by embedding similarity, not keywords
Hybrid searchCombining keyword (BM25) and vector search. Beats either alone
RerankingRe-scoring retrieved passages for relevance before the model sees them
Contextual retrievalAdding LLM-generated context to each chunk before embedding. Cuts miss rate ~49%
HallucinationA fluent, confident answer that is not supported by the source
GroundingForcing an answer to reference retrieved evidence

Agents & MCP

AgentAn LLM that plans, uses tools, and takes multiple steps toward a goal
Tool useLetting the model call functions, APIs, or databases
MCPModel Context Protocol. An open standard for connecting models to tools and data
OrchestratorA lead model that decomposes a task and delegates to workers
Agent harnessThe infrastructure around a model that lets it run long tasks reliably
Human-in-the-loopA required human approval step inside an automated flow
ReActReason, act (call a tool), observe, repeat. A common agent loop pattern
Function callingA model's ability to output a structured request to invoke a specific function
QR-04

AI Glossary Quick Card

Training & tuning, evaluation & cost, risk & governance.
Last verified 2026-07-09

Training & Tuning

Pre-trainingThe huge, expensive first pass where a model learns general knowledge
Fine-tuningFurther training on your data to bake in behaviour, tone, or format
LoRALow-Rank Adaptation. A cheap, popular fine-tuning method
RLHFReinforcement Learning from Human Feedback. Aligning outputs with human preferences
DistillationTraining a small model to imitate a larger one. Runs cheaper
Base vs InstructBase predicts text; Instruct follows instructions. Almost every product uses Instruct

Evaluation & Cost

EvalA test that scores model output against expected results
Golden datasetA trusted set of inputs and correct outputs. Your source of truth for evals
LLM-as-judgeUsing a model to grade another model's output. Cheap, imperfect
RAGASA framework of metrics for evaluating RAG systems (faithfulness, relevancy, etc.)
Batch APISubmitting many requests together for a discount, at higher latency
Input vs Output tokensBilled separately. Output usually costs 4x to 6x input
Cache readA cheaper price for input the provider has seen before
Frontier / Balanced / BudgetModel tiers by cost. Route work to the cheapest tier that clears the quality bar

Risk & Governance

GuardrailA check that blocks unsafe or off-policy output before it reaches the user
Prompt injectionHostile instructions hidden in user or retrieved content. Top production risk
JailbreakTechniques that trick a model into ignoring its safety training
BiasSystematic favouritism or unfairness in outputs, often traceable to training data
Data residencyWhere the data physically lives. A hard constraint for regulated industries
EU AI ActThe EU regulation classifying AI systems by risk and setting duties per class
NIST AI RMFA voluntary US framework for managing AI risk, widely adopted as a baseline
Model cardA document describing a model's capabilities, limits, and intended uses
Not here on purpose. Vendor brand names, model version numbers, and leaderboard rankings. Those move faster than a glossary can keep up, and belong in QR-02 (models) and QR-08 (frameworks), which are refreshed monthly.