Common AI Terms Explained: LLMs, RAG, MCP, Tokens, and Agents

Artificial intelligence has its own vocabulary, but many terms describe different parts of the same system. An LLM is a model; a prompt is input to that model; RAG supplies retrieved information; MCP can connect an AI application to tools and data; and an agent uses a model within a workflow that can take actions. Understanding those boundaries makes product descriptions and business proposals much easier to evaluate.

AI, machine learning, and generative AI

Artificial intelligence (AI) is the broadest term. It covers computer systems designed to perform tasks associated with abilities such as perception, prediction, language use, planning, or decision support. Not every AI system is a chatbot, and not every automated rule is meaningfully AI.

Machine learning is a way of building AI systems by learning patterns from data instead of writing a rule for every possible case. Fraud detection, recommendation systems, and image classification are familiar examples.

Generative AI is a category of models that creates new output—such as text, images, audio, video, or code—based on patterns learned during training and the input supplied at use time. The 2017 “Attention Is All You Need” paper introduced the Transformer architecture that became influential in modern language models, but generative AI is broader than one architecture or one model family.

Model, foundation model, and large language model

What is a model?

A model is a learned mathematical system that maps input to output. Different models may classify an image, transcribe speech, forecast demand, or generate text. The surrounding product—its interface, search, tools, safety controls, and storage—is not the model itself.

What is a foundation model?

A foundation model is trained broadly enough to be adapted to many downstream tasks. A provider may expose one through an API, while another company builds a customer-support assistant, coding tool, or document analyzer around it. Foundation models can be language-only or multimodal.

What is an LLM?

A large language model (LLM) is trained to model sequences of tokens and generate or transform language. It may answer questions, summarize, translate, extract structured data, or write code. Many current LLMs also accept images or audio, so “language model” does not necessarily mean text-only.

An LLM does not contain a searchable copy of every training document, and it does not independently verify every statement it produces. It generates output from learned patterns plus the context available during the request. For current model families and their practical differences, see TipsMake’s guide to leading large language models.

Prompt, response, and instruction hierarchy

A prompt is the input given to an AI system. It can include a question, task, examples, constraints, attached files, retrieved passages, and tool results—not just a single sentence typed by a user.

Many AI applications distinguish between instructions supplied by the system or developer and the user’s request. The exact hierarchy depends on the platform, but higher-priority instructions generally define behavior and boundaries while the user supplies the immediate task.

A useful prompt states the goal, relevant context, desired format, and important constraints. It should also say what the model should do when information is missing. A detailed prompt can improve the result, but it cannot guarantee accuracy.

What is a token?

A token is a unit used to represent input and output. A token may be a whole short word, part of a longer word, punctuation, whitespace, or another fragment. The split varies by model and tokenizer, so “one token equals one word” is only a rough shortcut.

Token counts matter because providers often use them to measure context limits, API usage, and cost. Images, audio, cached input, and reasoning may be measured differently, depending on the service.

What is a context window?

The context window is the amount of information a model can consider in one request or session step. It can include instructions, the conversation, documents, retrieved passages, tool outputs, and the response being generated.

A larger context window does not guarantee that every detail will be used correctly. Long or contradictory context can hide important instructions. Put essential requirements close to the task, remove irrelevant material, and use search or retrieval to supply the most relevant passages. Context limits vary by model and can change, so check the provider’s current documentation rather than relying on a fixed number.

What does temperature mean?

Temperature is a generation setting that influences how strongly a model favors more likely token choices. Lower settings usually make sampling more constrained; higher settings can increase variation. It is a numeric parameter, not a physical temperature, so it is not measured in degrees.

Temperature does not turn a model from “creative” into “accurate,” and a low value does not prevent factual errors. Some current reasoning models or products do not expose the setting, and implementations may behave differently. For extraction, classification, or repeatable formatting, use constrained output and validation in addition to any sampling setting.

RAG, embeddings, and fine-tuning

What is RAG?

Retrieval-augmented generation (RAG) retrieves relevant information from a chosen source and places it in the model’s context before generating an answer. A company assistant might retrieve sections from current policies, product documentation, or approved knowledge-base articles and cite them in its response.

RAG does not automatically make an answer correct. Quality depends on document freshness, access controls, chunking, search quality, the prompt, and whether the system requires the model to ground claims in retrieved evidence.

What are embeddings?

Embeddings are numeric representations used to compare semantic similarity. RAG systems often use them to find passages related to a query, although keyword search and reranking can also be part of retrieval. An embedding database is an index for finding content; it is not a replacement for the original documents or their permissions.

What is fine-tuning?

Fine-tuning continues training a model on curated examples to influence behavior or performance for a task. It can help with consistent formats, domain-specific patterns, or specialized classifications. It is usually a poor way to keep frequently changing facts current.

Use RAG when the system needs up-to-date or source-specific knowledge. Consider fine-tuning when you have representative training and evaluation data and need repeatable behavior that prompting alone does not deliver. Some systems use both.

What is alignment?

Alignment concerns whether an AI system behaves in accordance with intended goals, instructions, and human values. In product work, the term may include instruction following, refusal behavior, safety policies, and avoiding unintended actions.

Alignment is not a one-time switch. Developers use training, policies, permissions, monitoring, and evaluations to reduce unwanted behavior. A prompt such as “organize my inbox but never delete messages” is a constraint, but an application should also enforce that boundary through tool permissions rather than relying on text alone.

What is MCP?

Model Context Protocol (MCP) is an open standard for connecting AI applications to external systems. The official MCP documentation describes a host application that creates clients to communicate with MCP servers. Servers can expose resources, prompts, or tools.

• Host: the AI application the user interacts with.
• Client: the protocol component that maintains a connection to a server.
• Server: a program that exposes specific data or capabilities.

MCP is the connection standard, not the model, the phone, or the business application itself. It also does not make every integration trustworthy. A tool may read files, call APIs, or modify records, so users and administrators still need authentication, least-privilege access, clear confirmation for consequential actions, and audit logs.

What are AI agents and agentic AI?

An AI agent combines a model with instructions, tools, state, and a control loop so it can perform multiple steps toward a goal. Depending on its permissions, it might search records, create a draft, call an API, check the result, and ask the user for approval before a final action.

Agentic AI is the broader design approach or category of systems that can plan or act with some degree of autonomy. There is no single universal threshold that turns a chatbot into an agent, so product claims should be judged by actual capabilities and controls. TipsMake’s guide to AI agent frameworks shows how developers combine models, tools, memory, and workflows.

Autonomy raises risk as well as capability. Sending an email, entering an invoice, or changing a database is different from drafting text. Good systems restrict tools, validate inputs, show what will happen, request confirmation for consequential actions, and provide a record of the operation. The distinction is especially useful when evaluating computer-use agents that interact with software interfaces.

How the pieces fit together

Term	Role in a typical AI application	What it is not
LLM	Generates or transforms language from context	The complete chatbot or agent product
Prompt	Supplies a task, context, examples, and constraints	A guarantee of correct output
Token	Represents units of model input and output	Always a whole word
Context window	Limits information available in a model interaction	Permanent memory
RAG	Retrieves relevant source material for the model	Automatic fact verification
Fine-tuning	Adapts model behavior using curated examples	A live knowledge database
MCP	Standardizes connections to tools and data sources	An AI model
Agent	Uses a model, tools, state, and workflow to take steps	Necessarily fully autonomous

When evaluating an AI feature, ask four concrete questions: Which model is used? What information enters its context? Which tools can it call? What verifies or approves the result? Those answers are more informative than a broad “AI-powered” label.