AI Glossary

An AI that can plan steps and use tools (like search or spreadsheets) to complete tasks, not just answer a single prompt.

How software systems connect and share data. Lets apps talk to each other and to AI models.

A standardized test used to compare model performance (e.g., SWE-bench for coding).

Unfair patterns learned from data that can cause skewed or discriminatory outputs.

The maximum amount of text (in tokens) a model can consider at once. Larger windows let the AI read longer documents or conversations.

Splitting long documents into smaller pieces so they fit within the context window.

The step-by-step reasoning behind an answer. Often hidden or summarized for safety.

A numerical representation of text that captures its meaning. Used to compare similarity between pieces of text.

Measuring how well an AI performs using tests, examples, or benchmarks.

Training an existing model further on your data so it better matches your style, terminology, or tasks.

When an AI is given a few examples in the prompt to show it how to perform a task.

Connecting answers to trusted sources or documents so that responses are verifiable and accurate.

Controls that limit an AI’s behavior to stay safe and on-policy (filters, rules, constraints).

The correct answer or trusted source used to check an AI’s output.

When an AI makes up facts or gives an incorrect answer with confidence.

Tokens you send to the AI (your prompt and any context). Billed separately from outputs by many providers.

When the trained model runs to produce an answer. This is the “use” phase after training.

An AI system trained on large amounts of text that can understand and generate language (e.g., ChatGPT, Claude, Gemini).

How long the AI takes to respond. Lower latency feels faster and more interactive.

The “brain” of the AI—the learned system that produces answers based on patterns in data.

AI that can work with more than one type of input, such as text, images, audio, or video.

Information the AI or app keeps across interactions to personalize or continue a task.

Tokens the AI generates in its response. Often billed at a different rate than inputs.

A model whose code and/or weights are publicly available to use and modify (e.g., Llama, Mistral).

The instruction or question you give to an AI. Clear prompts lead to better results.

The practice of writing prompts in a structured way to guide the AI toward better, more reliable answers.

The internal values a model learns during training. More parameters can mean more capability, but also higher cost.

A closed, vendor-controlled model accessed via an API (e.g., OpenAI, Anthropic, Google Gemini).

Most providers charge based on tokens used for inputs and outputs.

A technique where the AI looks up relevant documents first and then uses them to answer your question, reducing hallucinations.

The maximum number of requests or tokens allowed in a given time period by a provider.

Hidden instructions that set the AI’s role and behavior (e.g., “You are a helpful assistant”).

Practices and controls to prevent harmful outputs and misuse, and to protect users and data.

A small chunk of text that AI models read and write. Roughly four characters of English on average. Pricing and limits are based on tokens.

A modern AI architecture that powers most LLMs. It handles long text efficiently by focusing on important parts (attention).

A setting that controls how creative or cautious the AI is. Higher values are more creative, lower values are more consistent.

A setting that limits the AI to the most likely words whose probabilities add up to p (e.g., 0.9), balancing variety and quality.

The AI’s ability to call external tools or functions, like searching the web or looking up data from your systems.

Teaching the model by showing it many examples so it learns patterns and can generalize to new inputs.

A database optimized for storing and searching embeddings to find the most similar items quickly.

When an AI solves a task without seeing any examples—only the instructions.