Learning Hub / AI glossary
PLAIN ENGLISH, USEFUL CONNECTIONS

A shared language for AI.

30 terms you will meet across the learning routes. Understand the idea, then see where it fits in practice.

Abstention

Choosing not to provide an unsupported answer or take an unjustified action.

In practice: State which information is missing and how it can be checked.

Explore: Confidence & uncertainty →

Agent

A system that chooses steps and uses tools.

In practice: Ask which actions it may perform without review.

Explore: AI agents & permissions →

Baseline

The reference result used to compare a change.

In practice: Keep the old workflow result alongside the proposed version.

Explore: Evaluation & risk →

Blind review

Judging outputs without seeing information that may bias the judgment, such as the model name.

In practice: Rotate answer order and use the same rubric.

Explore: Compare AI models fairly →

Calibration

How closely confidence estimates align with observed correctness across assessed predictions.

In practice: Do not treat a confidence percentage in one answer as proof of accuracy.

Explore: Confidence & uncertainty →

Chunk

A smaller passage prepared for retrieval.

In practice: Keep the title and version with a handbook passage.

Explore: RAG explained →

Citation

A reference identifying a supporting source.

In practice: Open the source and check the exact claim.

Explore: Verify AI outputs →

Context

The information available during a model interaction.

In practice: A document not supplied or retrieved may be unavailable.

Explore: Better prompts →

Context window

The bounded token capacity of a model interaction.

In practice: Check how your application handles long conversations.

Explore: AI foundations →

Data minimization

Using only the data needed for the task.

In practice: Remove email addresses before analyzing fictional feedback themes.

Explore: AI data privacy →

Embedding

A numerical representation used to compare meaning.

In practice: Similarity helps find candidates; it does not verify facts.

Explore: RAG explained →

Evaluation

Assessing performance against defined tasks and criteria.

In practice: Include cases where asking for help is correct.

Explore: Evaluation & risk →

Fine-tuning

Additional training that adjusts an existing model.

In practice: Distinguish changing model behavior from retrieving current documents.

Explore: Prompting, RAG & fine-tuning →

Grounding

Connecting an answer to supplied or retrieved evidence.

In practice: Check whether the selected passage actually supports the sentence.

Explore: Verify AI outputs →

Hallucination

Generated content that is false or unsupported.

In practice: A plausible but invented deadline needs correction.

Explore: AI foundations →

Held-out set

Evaluation examples kept separate from the examples used to train or tune a system.

In practice: Use fresh questions to check whether an improvement generalizes.

Explore: Compare AI models fairly →

Human review

A person checks an output or proposed action.

In practice: Give the reviewer evidence and authority to reject it.

Explore: Reliable AI workflows →

Inference

Using a trained model to produce an output.

In practice: A response is a model use, not necessarily new training.

Explore: AI foundations →

Latency

The elapsed time between a request and a defined response point.

In practice: Specify whether you measure the first token or the complete usable response.

Explore: Measure workflow value →

Least privilege

Granting only the access a task requires.

In practice: A drafting assistant may need no publishing permission.

Explore: AI agents & permissions →

Model

A learned system for predictions or generated outputs.

In practice: The model and its surrounding application have different roles.

Explore: AI foundations →

Prompt

Instructions and input supplied to guide a response.

In practice: Name the task, evidence, format, and quality checks.

Explore: Better prompts →

Prompt injection

Untrusted content attempts to redirect an AI system.

In practice: A retrieved page cannot grant permission to publish.

Explore: AI agents & permissions →

RAG

Retrieval-augmented generation: answering with selected external information.

In practice: Inspect retrieved passages before blaming the final response.

Explore: RAG explained →

Regression

A previously acceptable behavior worsens after a change.

In practice: Repeat the same cases when revising a prompt.

Explore: Evaluation & risk →

Rubric

Explicit criteria for judging a result.

In practice: Score factual support separately from polished writing.

Explore: Evaluation & risk →

Token

A unit of input or output representation.

In practice: A word can consist of several text tokens.

Explore: AI foundations →

Tool call

A request to execute an application function.

In practice: Check the destination and arguments before allowing a send.

Explore: AI agents & permissions →

Trial

One attempt at a defined evaluation task.

In practice: Repeat a missing-evidence question to inspect inconsistent behavior.

Explore: Compare AI models fairly →

Workflow

A defined sequence of steps and decisions.

In practice: Write what happens when an input is missing.

Explore: Reliable AI workflows →

Sources & further reading

These concise explanations use the terms as they appear in this hub. Definitions can vary across research fields and products.

Google Machine Learning Glossary ↗

Anthropic: Building effective agents ↗

OWASP: LLM application risks ↗

Choose a learning route →