Glossary of AI and exam terms

Plain-language meanings of the 61 terms used across the certifications on Rejelsa Prep.

Core AI ideas

Artificial intelligence (AI)
Software that performs tasks which normally need human intelligence, such as understanding language, recognising images or making predictions.
Machine learning
A way of building AI in which a system learns patterns from data instead of following rules written by hand.
Generative AI
AI that creates new content, such as text, images, audio or code, in response to a request.
Large language model (LLM)
A model trained on very large amounts of text that can read and write language. It works by predicting what text should come next.
Small language model
A more compact language model that is faster and cheaper to run, suited to simpler or narrower tasks.
Foundation model
A large, general-purpose model that can be adapted to many different tasks without being rebuilt for each one.
Multimodal model
A model that can work with more than one kind of input or output, for example text together with images or audio.
Token
A small unit of text, roughly a word or part of a word. Models read and write in tokens, and usage is measured and billed in them.
Context window
The maximum amount of text, counted in tokens, that a model can take into account in one request, including its own reply.
Training data
The examples a model learns from. The quality and coverage of this data shape what the model can do and where it is weak.
Fine-tuning
Giving an existing model additional training on specific data so that it performs better on a specialised task.
Inference
Using a trained model to produce an output, as opposed to training it.

Working with models

Prompt
The instruction or question given to a model. Clear prompts state the goal, the context and the form the answer should take.
System prompt
Standing instructions that set a model's role, tone and limits for a whole conversation, separate from the user's individual messages.
Prompt engineering
The practice of writing and refining prompts to get more accurate and useful responses.
Few-shot prompting
Including a few examples of the desired output in a prompt so the model can follow the pattern.
Temperature
A setting that controls how varied a model's output is. Lower values give more predictable answers, higher values more variety.
Grounding
Connecting a model's answer to specific, trusted information, such as company documents, so that it is based on facts and not only on training.
Retrieval-augmented generation (RAG)
A pattern in which relevant content is first retrieved from a knowledge source and then given to the model to answer from.
Embedding
A list of numbers that represents the meaning of a piece of text, so that similar meanings can be found by comparing numbers.
Chunking
Splitting documents into smaller passages before indexing them, so that retrieval can return the most relevant part.
Structured output
A model response in a fixed format, such as JSON that follows a schema, so that other software can use it reliably.
Prompt caching
Reusing an unchanged opening section of a prompt across requests to reduce cost and waiting time.

Agents and tools

Agent
An AI system that decides its own next steps and uses tools to complete a task, instead of only answering a single question.
Tool use (function calling)
A model asking the surrounding application to run a defined function, such as looking up an order, and then using the result.
Model Context Protocol (MCP)
An open standard for connecting AI applications to external tools and data sources in a consistent way.
Orchestration
Coordinating several steps, models or agents so that they work together on one task.
Subagent
An agent given one part of a larger task by a coordinating agent. It works in its own context and reports back.
Human in the loop
A design in which a person reviews or approves an AI system's work at key points, especially before actions that are hard to undo.
Copilot
Microsoft's name for its AI assistants, which work alongside a person inside products such as Word, Excel, Outlook and Teams.
Connector
A link that lets an AI assistant read from, or act in, another system such as a file store or ticketing tool.

Quality and safety

Hallucination (fabrication)
A response that sounds confident and plausible but is false or unsupported. Checking against sources is the main defence.
Groundedness
A measure of whether the claims in a response are supported by the source material it was given.
Evaluation
Testing an AI system against a set of examples with defined criteria, to measure quality and catch regressions.
Prompt injection
An attack in which text hidden in a document, web page or message tries to make an AI system follow someone else's instructions.
Guardrails
Controls that keep an AI system within safe limits, such as input checks, output filters and restricted tool permissions.
Responsible AI
Principles for building and using AI well: fairness, reliability and safety, privacy and security, inclusiveness, transparency and accountability.
Bias
Unfair differences in how a system treats people or groups, often inherited from patterns in its training data.
Data loss prevention (DLP)
Policies that detect sensitive information, such as card numbers, and stop it being shared inappropriately.
Sensitivity label
A tag applied to a file or email that classifies how sensitive it is and can add protection such as encryption.
Zero Trust
A security approach built on three ideas: verify every request explicitly, give the least access needed, and assume a breach is possible.
Least privilege
Giving each person or system only the access it needs for its job, and no more.

Exam terms

Exam outline
The official list of topics an exam covers and how much each one counts. Also called a blueprint, exam guide or study guide.
Exam domain
One of the main topic areas in an exam outline. Each domain has a weight showing its share of the exam.
Scaled score
A score converted to a fixed range, such as 100 to 1000, so that results are comparable between different versions of an exam. It is not a simple percentage.
Pass mark
The score needed to pass. It varies by provider: Claude exams pass at 720 of 1000, Microsoft at 700 of 1000, AWS between 700 and 750 of 1000, IAPP AIGP at 300 of 500 and CompTIA SecAI+ at 600 of 900. Some exams use a percentage instead, such as 85% for PSM I and 65% for ITIL 4 Foundation, and Google, NVIDIA and PMI report only pass or fail.
Proctored exam
An exam taken under supervision, either at a test centre or online with a remote invigilator.
Mock exam
A practice exam taken under realistic conditions, with the same pace and time pressure as the real one.
Multiple-response question
A question with more than one correct option, where you must select all the correct ones. The question states how many to choose.

Project, agile and service terms

Sprint
A fixed period of one month or less in which a Scrum Team creates a usable, valuable increment of the product.
Product Backlog
The single, ordered list of everything that might be needed to improve a product. The Product Owner is accountable for it.
Definition of Done
A shared description of the quality an increment must meet before it counts as complete.
Velocity
The amount of work a team usually completes in an iteration. It helps the team plan, but it should not be used to compare teams.
Critical path
The longest chain of dependent activities in a schedule. It sets the shortest possible project duration.
Earned value
A way of measuring project performance by comparing the value of work done with the planned work and the actual cost.
Hybrid approach
Combining predictive and agile methods in one project, using each where it fits best.
Incident and problem
In ITIL, an incident is an unplanned interruption to a service; a problem is the cause, or potential cause, of one or more incidents.
Change enablement
The ITIL practice that assesses risk and authorises changes so that as many as possible succeed.
Value stream
The series of steps an organisation takes to create and deliver products and services to customers.