Artificial intelligence is changing how people learn, work, build software and make decisions. But understanding AI can be difficult when every article introduces unfamiliar terms such as large language models (LLMs), generative AI, embeddings, retrieval-augmented generation (RAG) and agentic AI.
This AI glossary explains more than 50 essential AI terms in plain English, with practical examples to help you understand what they mean and why they matter.
Whether you are a student exploring artificial intelligence, a professional evaluating AI tools, a business owner considering automation, or a developer building AI applications, use this guide as a reference whenever a technical term feels unfamiliar.
Quick answer: Artificial intelligence (AI) is the broader field of creating systems that perform tasks associated with human intelligence. Machine learning is one approach within AI, deep learning is a subset of machine learning, and generative AI focuses on creating new content. Large language models power many text-based AI applications, while techniques such as RAG and systems such as AI agents extend how those models can access information and perform tasks.
AI Fundamentals
1. Artificial Intelligence (AI)
Artificial intelligence is a field of computing focused on creating systems that perform tasks such as recognizing patterns, understanding language, making predictions or generating content.
Example: An AI system that identifies suspicious transactions can help a financial institution flag activity for further review.
2. AI Model
An AI model is a trained computational system that uses patterns learned from data to produce predictions, classifications or generated outputs.
Example: A language model generates a draft email from instructions provided by a user.
Read more: What Is an AI Model?
3. AI System
An AI system combines a model with the surrounding software, data, interfaces, rules and operational processes needed to perform a task.
Example: A customer-support application may combine a language model, a company knowledge base, a chat interface and escalation rules.
Remember: An AI model is one component; an AI system is the complete working arrangement.
Read more: What Is an AI System?
4. Algorithm
An algorithm is a defined sequence of computational steps used to solve a problem or perform a task.
Example: A search algorithm ranks results according to a set of criteria. A machine-learning model may learn patterns from data, while an algorithm describes the computational procedure used to process it.
5. Dataset
A dataset is a collection of data used for analysis, model training, evaluation or other computational tasks.
Example: A collection of labeled photographs may be used to train a model to recognize different types of objects.
6. Training
Training is the process through which a machine-learning model adjusts its internal parameters using data and an optimization procedure.
Example: During language-model training, the model learns statistical patterns from large collections of text.
7. Inference
Inference is the process of using a trained model to produce an output from new input.
Example: When you ask an AI assistant to summarize a report, the model performs inference to generate the summary.
8. Parameters
Parameters are internal numerical values learned during model training. They influence how a model processes information and generates outputs.
Example: A language model’s learned parameters help determine which tokens are likely to follow a given sequence.
A larger parameter count does not automatically mean a model will be better for every task.
Machine Learning and Deep Learning
9. Machine Learning (ML)
Machine learning is a branch of AI in which systems learn patterns from data to make predictions or decisions without requiring every rule to be explicitly programmed.
Example: A model learns from historical transactions to estimate whether a new transaction may be fraudulent.
10. Deep Learning
Deep learning is a type of machine learning that uses neural networks with multiple processing layers to learn complex patterns.
Example: Deep-learning models are used in speech recognition, image analysis and many modern language applications.
11. Neural Network
A neural network is a computational model composed of interconnected units that transform input information through layers of calculations.
Example: A neural network can learn patterns in images that help distinguish a cat from a dog.
The name is inspired by biological neural networks, but artificial neural networks are mathematical systems rather than replicas of the human brain.
12. Transformer
A transformer is a neural-network architecture that uses attention mechanisms to process relationships between elements in a sequence.
Transformers underpin many modern language models and are also used in image, audio and multimodal applications.
Example: In a sentence, attention helps a model relate a pronoun to a relevant noun elsewhere in the text.
13. Supervised Learning
Supervised learning trains a model using examples that include inputs and corresponding target labels or values.
Example: A model learns to classify emails as spam or not spam from labeled training examples.
14. Unsupervised Learning
Unsupervised learning uses data without predefined target labels to identify patterns, structures or groupings.
Example: A system groups customers according to similarities in their purchasing behavior.
15. Reinforcement Learning
Reinforcement learning trains an agent to make decisions through interactions with an environment and feedback expressed as rewards or penalties.
Example: An agent learns a strategy in a simulated environment by receiving feedback on its actions.
16. Fine-Tuning
Fine-tuning adapts an already trained model using additional training data to improve its behavior for a particular task or domain.
Example: A company may fine-tune a language model using carefully prepared examples to improve its performance on a specialized classification task.
Fine-tuning is not automatically the best way to add new factual information. For knowledge that changes frequently, retrieval-based approaches may be more suitable.
17. Overfitting
Overfitting occurs when a model learns the details of its training data so closely that it performs poorly on new, unseen examples.
Example: A model may perform extremely well on practice questions but struggle with different questions that test the same underlying skill.
18. Evaluation Dataset
An evaluation dataset is a collection of examples used to assess how a model performs on specified tasks.
Example: A team may test a chatbot against representative customer questions with expected answers and predefined scoring criteria.
A good evaluation dataset should reflect the real tasks and conditions in which the system will be used.
Generative AI and Large Language Models
19. Generative AI
Generative AI refers to AI systems designed to produce new content, such as text, images, audio, video or code.
Example: An image-generation model creates an illustration from a written description.
Generative AI can produce useful material, but its outputs still require appropriate review.
20. Foundation Model
A foundation model is a model trained on broad data that can be adapted or used across multiple tasks.
Some foundation models are multimodal, meaning they can process or generate more than one type of data.
Example: A foundation model may support summarization, question answering and document analysis through different applications.
21. Large Language Model (LLM)
A large language model is a model trained to process and generate language. Many LLMs generate text by predicting tokens based on the preceding context.
LLMs power chatbots, writing assistants, coding tools, summarization systems and many other applications.
Example: An LLM can explain a technical concept in simpler language or help draft a report.
Read more: What Is an LLM (Large Language Model)?
22. Multimodal AI
Multimodal AI processes or generates more than one type of data, such as text, images, audio or video.
Example: A user uploads a chart and asks an AI assistant to explain its contents in text.
Multimodal capability varies by model; not every AI assistant supports every data type.
23. Token
A token is a unit of data processed by a language model. For text, a token may represent a whole word, part of a word, punctuation or another text segment.
Tokens matter because models process text in token sequences, and many model APIs calculate usage costs using input and output tokens.
Example: The sentence “AI can help” is converted into tokens before a language model processes it. The exact tokenization depends on the model.
Read more: What Is a Token in AI?
24. Context Window
A context window is the amount of input and conversation context a model can process within a particular request.
A larger context window can allow a model to consider more text, but it does not guarantee that the model will accurately use every detail.
Example: A model with a suitable context window may analyze a long document without requiring the user to split it into many separate prompts.
25. Hallucination
An AI hallucination occurs when a model generates information that is false, unsupported or inconsistent with the available evidence.
Example: An AI assistant invents a research paper title or attributes a statement to a source that does not contain it.
Verify important claims against reliable sources, especially when the output concerns health, law, finance or other consequential decisions.
26. Reasoning Model
A reasoning model is designed or trained to perform tasks that benefit from more extensive problem-solving or multi-step analysis.
Example: A reasoning-focused model may help work through a complex mathematics problem or compare several technical constraints.
The label does not guarantee correctness. Performance depends on the task, model, available information and evaluation method.
27. Open-Weight Model
An open-weight model makes its trained parameters available for others to download or run, subject to the model’s license and usage conditions.
Open weights do not necessarily mean the training data, complete training process or every associated component is openly available.
Example: A developer downloads an open-weight model and runs it in a controlled environment.
28. Model Quantization
Quantization reduces the numerical precision used to represent model parameters or computations, often to reduce memory requirements and improve inference efficiency.
Example: A quantized model may require less memory to run locally than a higher-precision version.
The trade-off is that quantization can affect model quality, speed or hardware compatibility depending on the method.
Prompts and Model Responses
29. Prompt
A prompt is the input provided to an AI model to guide the response or task.
It may contain instructions, questions, context, examples or supported media.
Example: “Summarize this report in five bullet points for a non-technical audience” is a prompt.
30. Prompt Engineering
Prompt engineering is the practice of designing and refining prompts to improve the relevance, format or quality of model outputs.
Example: Instead of asking an AI to “write about cybersecurity,” specify the audience, topic, desired length and evidence requirements.
Clear prompts can improve results, but they cannot guarantee factual accuracy.
31. Zero-Shot Prompting
Zero-shot prompting asks a model to perform a task without providing task-specific examples in the prompt.
Example: “Classify this review as positive, neutral or negative.”
32. Few-Shot Prompting
Few-shot prompting provides a small number of examples to show a model the desired pattern or output format.
Example: A prompt includes three examples of product reviews and their categories before asking the model to classify a fourth.
33. System Instructions
System instructions define important behavior, role, constraints or rules for an AI interaction, depending on the platform and model interface.
Example: An application may instruct an assistant to answer only from an approved knowledge base and to say when the evidence is insufficient.
System instructions do not eliminate all errors or override every limitation of the model.
34. Structured Output
Structured output is information generated in a predefined format, such as JSON or a schema-defined set of fields.
Example: A model extracts a product name, price and category into separate fields rather than returning a free-form paragraph.
Structured output makes integration with software easier, but the resulting data still needs validation.
AI Agents and Automation
35. AI Agent
An AI agent is a software system that uses a model, instructions and available tools to pursue a goal through one or more steps.
Unlike a simple text-generation interaction, an agent may decide which tool to use, inspect the result and continue working toward its objective.
Example: A research agent might search approved sources, collect relevant documents and prepare a draft summary for human review.
The level of autonomy varies. Some agents require approval before acting, while others can perform a defined set of actions without continuous intervention.
36. Agentic AI
Agentic AI describes systems designed to pursue goals through planning, decisions and actions, often with some degree of autonomy.
The term is used inconsistently across the industry, so it is important to examine what a specific product actually does.
Example: A system that plans a multi-step research task and invokes tools to complete it may be described as agentic.
37. AI Workflow
An AI workflow is a sequence of steps that combines AI operations with other software or business processes.
Example: A workflow receives a support ticket, classifies the issue, drafts a response and routes it to an employee for approval.
A workflow can be deterministic, AI-driven or a combination of both.
38. Tool Calling
Tool calling allows a model to request an external function, API or service to perform a specific operation.
Example: An assistant calls a calculator tool to perform a calculation rather than relying solely on generated text.
The application typically executes the approved tool call and returns its result to the model.
39. Function Calling
Function calling is a structured mechanism through which a model specifies a function to invoke and supplies arguments in an expected format.
It is closely related to tool calling and is frequently used to connect language models to external systems.
Example: A model produces structured arguments for a weather lookup function, and the application uses those arguments to request weather data.
40. Agent Orchestration
Agent orchestration coordinates the steps, tools, models or agents involved in a larger task.
Example: A research process may coordinate source discovery, document extraction, fact checking and human approval.
Orchestration determines how the parts work together and what happens when a step fails.
41. Human-in-the-Loop (HITL)
Human-in-the-loop describes a process in which people review, approve, correct or guide an AI system at important stages.
Example: An AI system drafts a customer email, but an employee approves it before sending.
Human review is particularly important when actions are consequential, difficult to reverse or dependent on context the system may not understand.
RAG, Embeddings and AI Search
42. Retrieval-Augmented Generation (RAG)
Retrieval-augmented generation combines information retrieval with generative AI. Before generating an answer, the system retrieves relevant information from an external source and supplies it to the model as context.
Example: An internal company assistant searches approved policy documents and uses relevant passages to answer an employee’s question.
RAG can improve grounding and access to current or specialized information, but it does not guarantee accuracy. Poor retrieval, outdated documents or incorrect interpretation can still produce wrong answers.
43. Embedding
An embedding is a numerical representation of data that captures useful relationships between items for a particular model or task.
Example: A semantic search system converts documents and user queries into vectors so it can identify passages related to a question even when they do not share exactly the same words.
Read more about related concepts in the QueryFeedz AI glossary.
44. Vector Database
A vector database stores and searches vector representations, often alongside metadata used to filter results.
Example: A document-search application stores embeddings of its knowledge base and retrieves passages that are semantically similar to a user’s query.
A vector database is one possible component of a RAG system; not every RAG implementation requires a dedicated vector database.
45. Semantic Search
Semantic search aims to find information based on meaning and context rather than relying only on exact keyword matches.
Example: A search for “I cannot access my account” may retrieve a password-reset guide even if the page uses different wording.
Semantic search can complement traditional keyword search, particularly when users describe a problem in natural language.
46. Vector Search
Vector search retrieves items by comparing vector representations according to a similarity or distance measure.
Example: A system compares the vector for a question with vectors representing stored documents and returns the closest matches.
Vector similarity is useful for retrieval, but a close vector match does not automatically mean a result is factually correct or relevant.
47. Grounding
Grounding connects a model’s response to specific information sources, such as supplied documents, databases or verified search results.
Example: A research assistant uses a company’s approved documentation to answer a question about its internal procedures.
Grounding can make an answer easier to verify, especially when the system provides traceable citations, but the retrieved evidence must still be checked.
48. Model Context Protocol (MCP)
Model Context Protocol (MCP) is an open protocol designed to standardize how AI applications connect with external tools and data sources through compatible servers and clients.
Example: An AI application may use an MCP server to access an approved document repository through a defined interface.
MCP is not itself an AI model or an autonomous agent. Actual capabilities and permissions depend on the implementation.
49. Retrieval
Retrieval is the process of finding relevant information from a collection of documents, records or other data sources.
Example: A RAG system retrieves the policy passages most relevant to an employee’s question before generating an answer.
Retrieval quality strongly influences the usefulness of the final response.
50. Reranking
Reranking reorders an initial set of search results to place the most relevant candidates higher in the list.
Example: A search system retrieves 30 potentially relevant passages and uses a reranker to identify the strongest candidates for the final answer.
Reranking can improve relevance, although its effectiveness depends on the model, query and retrieved material.
AI Evaluation, Reliability and Safety
51. AI Benchmark
An AI benchmark is a defined test or collection of tests used to compare model performance on particular tasks.
Example: A benchmark evaluates how well models answer questions, solve coding problems or process a specified dataset.
A benchmark score is useful only when its tasks, methodology and limitations are understood. A model that performs well on one benchmark may not be the best choice for a different application.
52. Evaluation Metrics
Evaluation metrics are measures used to assess a model or AI system against defined criteria.
Examples include accuracy, precision, recall, latency, task completion rate and cost per request.
Example: A document-classification system may be assessed using precision and recall, while an AI agent may also require evaluation of tool-call success and task completion.
Choose metrics that reflect the real objective rather than relying on one overall score.
53. Precision
Precision measures the proportion of predicted positive results that are actually positive.
For example, if a system flags 100 transactions and 80 are genuinely fraudulent, its precision is 80%, assuming those are the only flagged cases being evaluated.
54. Recall
Recall measures the proportion of all actual positive cases that a system successfully identifies.
For example, if there are 100 fraudulent transactions and a system identifies 90 of them, its recall is 90%.
Precision and recall often involve trade-offs. The right balance depends on the cost of false alarms and missed cases.
55. Latency
Latency is the time taken for a system to respond or complete an operation.
Example: A voice assistant with lower response latency can feel more conversational because users wait less time for a reply.
When evaluating AI systems, distinguish the time until the first output appears from the time required to complete the entire response.
56. AI Bias
AI bias refers to systematic patterns in a model’s outputs that may produce unfair or unbalanced results.
Bias can arise from training data, model design, evaluation choices or the context in which a system is deployed.
Example: A screening model may produce systematically different error rates for different groups.
Identifying and addressing bias requires appropriate testing, representative data and ongoing monitoring.
57. Responsible AI
Responsible AI refers to the practices used to develop and operate AI systems with attention to safety, fairness, privacy, transparency and accountability.
Example: A company establishes access controls, evaluates model errors and requires human approval for high-impact decisions.
Responsible AI is an ongoing process, not a one-time certification or a guarantee that a system cannot fail.
58. Explainable AI (XAI)
Explainable AI includes methods designed to help people understand aspects of how an AI system produces its outputs.
Example: A model may provide feature-importance information to help analysts understand which input variables influenced a prediction.
An explanation is not necessarily a complete description of a model’s internal computation, so it should be interpreted according to the method used.
59. Guardrails
Guardrails are technical or procedural controls intended to limit unsafe, incorrect or unauthorized behavior in an AI application.
Example: A support assistant may be restricted from accessing confidential records unless the user has the required permissions.
Guardrails reduce particular risks but do not eliminate every possible failure mode.
60. Prompt Injection
Prompt injection is an attack in which untrusted input attempts to manipulate an AI system into ignoring intended instructions or performing unauthorized actions.
Example: A malicious instruction hidden in a document tries to persuade a document-processing agent to reveal information it should not disclose.
Defenses can include separating trusted instructions from untrusted data, limiting tool permissions, validating actions and requiring approval for sensitive operations.
61. AI Red Teaming
AI red teaming is the practice of testing an AI system with adversarial or challenging inputs to identify weaknesses before or during deployment.
Example: A security team tests whether an assistant can be manipulated into exposing restricted information or misusing a connected tool.
Findings should inform practical mitigations, regression tests and monitoring.
How to Learn AI Terminology
You do not need to memorize every AI term to use artificial intelligence effectively. Focus first on the concepts that match what you want to do.
If you are new to AI
Start with artificial intelligence, machine learning, generative AI, AI models, LLMs, prompts, tokens and hallucinations.
These concepts help you understand what common AI applications can do and why their responses sometimes fail.
If you use AI tools at work
Learn about context windows, prompt engineering, structured output, grounding, human-in-the-loop processes and responsible AI.
These terms help you evaluate outputs, provide better instructions and identify when human review is necessary.
If you are building AI applications
Focus on embeddings, RAG, vector search, tool calling, MCP, AI agents, evaluation metrics, latency and prompt injection.
These concepts help you understand how models interact with external data, tools and operational systems.
If you evaluate AI models
Learn about benchmarks, evaluation datasets, precision, recall, inference, fine-tuning and quantization.
Understanding these terms helps you compare models according to actual requirements instead of relying only on model size or headline benchmark scores.
Frequently Asked Questions About AI Terminology
What is an AI glossary?
An AI glossary is a reference guide that explains artificial intelligence terms, acronyms and technical concepts. It helps beginners and professionals understand the terminology used in AI tools, research, product documentation and industry news.
What is the difference between AI, machine learning and deep learning?
AI is the broad field of creating systems that perform intelligent tasks. Machine learning is a subset of AI in which systems learn patterns from data. Deep learning is a subset of machine learning that uses multilayer neural networks.
What is the difference between generative AI and an LLM?
Generative AI is a broad category of systems that create content. A large language model is a type of model focused on language. LLMs can power text-generation applications, while generative AI also includes image, audio and video generation systems.
What is the difference between an AI agent and a chatbot?
A conventional chatbot primarily responds to user inputs. An AI agent may also plan steps, call tools and act toward a goal. Some chatbots include agent-like capabilities, so the distinction depends on the actual system rather than the product label.
What is the difference between RAG and fine-tuning?
RAG retrieves relevant information and provides it to a model as context at response time. Fine-tuning changes a model’s learned parameters through additional training. RAG is often useful when a system needs access to changing or specialized information; fine-tuning may be useful when a model needs to adapt to a task, format or behavior.
Can AI always provide accurate answers?
No. AI systems can generate incorrect information, miss context, retrieve irrelevant sources or make reasoning errors. Important outputs should be checked against reliable evidence, especially when decisions have significant consequences.
Why do AI terms change so quickly?
AI research, products and deployment practices evolve rapidly. New terminology appears as techniques develop, while existing terms may be used inconsistently by different vendors and researchers. Checking technical documentation and the date of a source helps clarify what a term means in a particular context.
Final Thoughts
Understanding AI terminology makes it easier to follow technology news, evaluate software and make informed decisions about using artificial intelligence.
Start with the fundamentals, then explore the concepts most relevant to your work. When you encounter an unfamiliar term, look beyond the definition: understand how the technique works, where it is useful and what its limitations are.
Explore more: Visit the QueryFeedz AI Glossary for individual explainers on AI models, LLMs, tokens, retrieval-augmented generation and other essential concepts.
Editorial note: AI terminology and technical implementations evolve. Definitions should be reviewed against current primary documentation before publication and during future updates.
