AI Chatbot Glossary: 40+ Terms Explained

How is this AI chatbot glossary organized?

This glossary groups terms by the layer they belong to. AI chatbot vocabulary falls into four layers: the model that generates text, the behavior you observe when you chat, the training that shapes the model, and the application that wraps the model for a job. Reading the layers in order builds a mental map, so a new term slots into place instead of floating loose.

Each entry gives a one-line definition, then a sentence of context that explains why the term matters when you pick or use a chatbot. Terms link together. A token feeds a context window, a context window bounds a prompt, and a prompt drives the output you read. Once you see the chain, the pricing tables and product pages stop reading like code.

Skim the layer that matches your question, then follow the cross-links to related terms.

What are the core model terms?

The model layer describes the engine under a chatbot. These terms show up in benchmark charts, model cards, and the fine print of an API plan. Learn them first, because every other layer builds on them.

• ▸Large language model (LLM): a neural network trained to predict text, the engine behind a modern chatbot such as ChatGPT or Claude.
• ▸Token: a chunk of text, such as a word or part of a word, that a model reads and writes. The word chatbot can split into two tokens.
• ▸Context window: the maximum number of tokens a model holds at once in a conversation, which sets how much of a document or chat it can weigh in one answer.
• ▸Parameters: the internal values a model adjusts during training. A higher count can mean more capability, though data quality and method matter as much as size.
• ▸Weights: the saved parameter values that define a trained model. To run a model is to load its weights and feed it tokens.
• ▸Multimodal: a model that handles more than text, such as images, audio, or video, in the same session.
• ▸Embedding: a list of numbers that captures the meaning of a piece of text, used to match a question against similar content.
• ▸Inference: the act of running a trained model to produce an answer, as opposed to training it. Your monthly bill pays for inference.

How tokens and context windows work together

Tokens are the unit a model counts, and the context window is the budget. If a model has a 200,000 token window, it can hold about 150,000 words of prompt and reply at once. When a chat runs past the window, the oldest text drops out, which is why a long conversation can lose an early instruction. A larger window lets a chatbot read a whole contract or codebase and answer without forgetting the start.

What do the behavior and quality terms mean?

The behavior layer covers what you observe when you chat. These terms explain why one answer sounds sharp and another drifts, and they name the settings you can adjust to steer output.

• ▸Prompt: the text you send to a chatbot, plus any hidden instructions the product adds before your words.
• ▸System prompt: the hidden instructions that set a chatbot's role, tone, and rules for a whole session.
• ▸Hallucination: a confident answer that states wrong information, produced because a model predicts plausible text rather than checked text.
• ▸Grounding: connecting a model to source data so answers stay tied to facts you can verify.
• ▸Temperature: a setting that controls how varied the output is. A low value gives steady, repeatable answers; a high value gives loose, inventive ones.
• ▸Reasoning model: a model that works through steps before it answers, which raises accuracy on math, logic, and code at the cost of speed.
• ▸Streaming: the way a reply appears word by word instead of all at once, so you can read as the model writes.
• ▸Latency: the delay between your message and the first word of the reply. Lower latency feels snappy; higher latency feels sluggish.
• ▸Chain of thought: the visible or hidden steps a reasoning model takes to break a problem into parts before it commits to an answer.

Why hallucinations happen and how to cut them

A hallucination is not a bug in the usual sense. A language model predicts the next token that fits the pattern, and a fluent guess can read as fact even when the source does not exist. The fix is grounding. When a chatbot pulls from a document or a live search and cites what it used, you can check the claim against the source. For anything that carries weight, such as a legal figure or a medical dose, treat the first answer as a draft and confirm it.

Grounding plus a citation is the strongest guard against a confident wrong answer.

What are the training and tuning terms?

The training layer names the stages that turn raw text into a helpful assistant. You will meet these terms in model announcements and in debates about safety and cost.

• ▸Pretraining: the first stage, where a model learns to predict text from a large body of data and picks up broad language and knowledge.
• ▸Fine-tuning: further training on curated examples to shape behavior for a task or a tone.
• ▸RLHF (reinforcement learning from human feedback): a method where people rate answers so the model learns which responses to prefer.
• ▸Alignment: the broader effort to steer a model toward useful, honest, and safe answers.
• ▸Instruction tuning: training that teaches a base model to follow requests, which turns a text predictor into an assistant.
• ▸Open weights: model files a provider releases so others can run, inspect, or tune the model on their own hardware.
• ▸Distillation: training a smaller model to copy a larger one, which lowers cost and speeds up inference.
• ▸Knowledge cutoff: the date after which a model saw no training data, which is why it needs a live source for recent facts.

Base model versus tuned assistant

A base model out of pretraining can predict text but does not follow orders well. Instruction tuning and RLHF turn that raw predictor into an assistant that answers questions, refuses harmful requests, and holds a helpful tone. When a provider ships a new chatbot, most of the visible personality comes from this tuning stage, not from the base model alone.

What are the application and business terms?

The application layer covers how a company wraps a model for a job. These terms show up in support software, sales demos, and enterprise contracts.

• ▸RAG (retrieval-augmented generation): a method where a chatbot pulls from documents and answers from that content with citations, which keeps replies tied to approved sources.
• ▸Agent: a chatbot that takes actions across steps and tools, not just replies. It can plan a task, use connected systems, and finish work such as booking a meeting.
• ▸Tool use: a model's ability to call outside functions, such as a search, a calculator, or a database, to complete a request.
• ▸Knowledge base: the approved content a business chatbot answers from, such as help articles and product docs.
• ▸Handoff: the moment a chatbot passes a conversation to a human agent when it reaches its limit.
• ▸Deflection: a support metric for the share of tickets a chatbot resolves without a human.
• ▸Answer engine: a search-first chatbot that returns cited answers, such as Perplexity.
• ▸Guardrails: the rules that block a chatbot from unsafe or off-topic replies.
• ▸MCP (Model Context Protocol): a standard that lets a chatbot connect to outside tools and data through a shared interface.
• ▸Vector database: a store of embeddings that a RAG system searches to find text close in meaning to a question.

How RAG and agents differ

RAG and agents solve different problems. RAG makes a chatbot accurate about your content by feeding it the right documents before it answers. An agent makes a chatbot capable of action by letting it call tools and chain steps. A support bot that quotes your help center uses RAG. A bot that reads a ticket, checks an order in your system, and issues a refund is acting as an agent. Many products blend both: the agent plans the steps, and a RAG step supplies the facts.

What do the safety, privacy, and access terms mean?

These terms decide whether a chatbot is safe for work data and who inside a company can use it. Read them before you paste sensitive text into any tool.

• ▸Training opt-out: a setting that stops a provider from using your inputs to train future models.
• ▸Data retention: how long a provider stores your chats before deletion.
• ▸SSO (single sign-on): a login control common in enterprise plans, which lets staff sign in with a company account.
• ▸SOC 2: an audit standard that shows a provider follows set controls for security and privacy.
• ▸PII (personal identifying information): data that can identify a person, such as a name, email, or account number.
• ▸Zero data retention: a mode where a provider processes your text and keeps no copy after the reply.
• ▸Role-based access: controls that limit which staff can see or change parts of a chatbot deployment.
• ▸Prompt injection: an attack where hidden text in a document or page tricks a chatbot into ignoring its rules.

What to check before you share sensitive text

1. 1Confirm the plan offers a training opt-out or keeps your text out of training by default.
2. 2Read the data retention window and delete chats that hold sensitive text.
3. 3Prefer a business or enterprise plan for client data, since consumer tiers can carry weaker guarantees.
4. 4Keep passwords, secrets, and full account numbers out of any consumer chatbot.
5. 5For a deployed bot, add guardrails and test it against prompt injection before launch.

A training opt-out and a short retention window are the two settings that matter most for work data.

What are the interface and integration terms?

These terms describe how you reach a chatbot and how it connects to other software. They matter when you compare a consumer app against a plan you build on.

• ▸API (application programming interface): an interface that lets software send prompts to a model and receive responses, which is how developers build a chatbot into an app.
• ▸Endpoint: the web address an API call reaches to run a model.
• ▸Rate limit: a cap on how many requests or tokens you can send in a window of time.
• ▸Plugin: an add-on that gives a chatbot a new skill, such as reading a live web page or running code.
• ▸Voice mode: an interface where you speak to a chatbot and it speaks back, built on speech-to-text and text-to-speech.
• ▸Widget: the chat box a company embeds on a website so visitors can ask questions in place.
• ▸Webhook: a callback that lets a chatbot or agent notify another system when an event happens.

Consumer app versus API access

The same model reaches you two ways. A consumer app such as the ChatGPT website gives a polished chat box, memory, and file upload with no setup. API access gives raw calls to the model so a developer can wire it into a product, set a system prompt, and control the flow. A small team that wants to draft and research picks the app. A team that wants a chatbot inside its own software picks the API and pays per token.

What are the cost and pricing terms?

Chatbot pricing runs on two models: a flat monthly seat or a per-token meter. Knowing the terms keeps a bill from surprising you.

• ▸Per-token pricing: a charge based on the tokens you send and receive, common for API access.
• ▸Input tokens and output tokens: the two halves of a request that a provider prices, often at different rates, with output priced higher.
• ▸Seat: a per-user monthly fee, common for consumer and team plans near twenty dollars a month.
• ▸Freemium: a free tier with limits that nudges heavy users toward a paid plan.
• ▸Prompt caching: a discount for reusing a fixed block of context, such as a long system prompt, across many calls.
• ▸Batch pricing: a lower rate for requests you can wait on, processed when the provider has spare capacity.

A worked example of a token bill

Suppose you build a support bot on an API. Each answer sends a 2,000 token system prompt, pulls 3,000 tokens of help articles through RAG, and returns a 500 token reply. That is 5,000 input tokens and 500 output tokens per answer. At a sample rate of three dollars per million input tokens and fifteen dollars per million output tokens, one answer costs about 2.3 cents. Ten thousand answers a month land near 230 dollars. Prompt caching on the fixed system prompt trims the input cost, which is why providers push it for high volume bots.

How do these terms connect in a live chatbot?

A single support question touches most of this glossary. Tracing one request shows how the layers link, so the terms stop feeling like a list and start feeling like a system.

1. 1You type a question into a widget on a store page. Your text becomes the prompt.
2. 2The product adds a system prompt that sets the bot's role and rules.
3. 3A RAG step turns your question into an embedding, searches a vector database, and pulls matching help articles.
4. 4The model reads the prompt plus the retrieved text, all inside its context window, and predicts a reply token by token.
5. 5The answer streams back with a citation, which grounds the claim in a source you can open.
6. 6If the bot cannot resolve the issue, it performs a handoff to a human agent, and the resolved share counts toward deflection.

Every term in that flow appears in the sections above. The prompt and system prompt come from the behavior layer, the embedding and context window from the model layer, RAG and handoff from the application layer. Once you can narrate a request end to end, you can read any product page and know what each feature does.

Which AI chatbot terms get confused most?

A few pairs trip up new users because the words sound close but mean different things. Sorting them out saves confusion on product pages and in sales calls.

The chatbot versus agent line matters most in 2026, because vendors label both as agents to signal capability. Ask a plain question: does the tool only answer, or does it act inside other software? If it books, edits, or buys, it is an agent. If it explains and drafts, it is a chatbot. The rest of the confusion clears once you place each term in its layer.