AI Glossary - Key Terms, in Plain Language
A free Synthreo glossary of the key AI terms in plain language: LLMs, tokens, context windows, RAG, agents, hallucinations, and more. No sign-in needed.
About this glossary
Section titled “About this glossary”The words below are the vocabulary of everyday AI conversations. These terms center on the generative AI and assistants most people work with day to day. Each definition is plain-language and vendor-neutral - it applies to those tools broadly, not just Synthreo’s. Pair it with the How to Speak AI guide, and when you are ready to learn by doing, the Synthreo Certified program takes you deeper.
Agent. An AI system defined by goal-directed action and orchestration: given an objective, it plans its own steps, calls tools or other systems, and works through a multi-step task with limited hand-holding. The line can blur - assistants can call tools too - but the shorthand holds: a plain chatbot mostly answers, an agent pursues a goal.
AI (Artificial Intelligence). Software that performs tasks normally associated with human intelligence, such as understanding language, recognizing patterns, or making decisions. In many current business conversations, “AI” usually means a large language model - though the term also covers vision, speech, recommendation, and forecasting systems.
Alignment. How well an AI system’s behavior matches what people actually want and intend. A well-aligned model is helpful, honest, and stays within safe boundaries.
API (Application Programming Interface). A standardized way for one piece of software to talk to another. An “AI API” lets your own applications send text to a model and get a response back, so AI can be built into products and workflows.
API key. A secret credential that identifies and authorizes you when your software calls an AI service. Treat it like a password - anyone who has it can spend on your account.
Chatbot / Assistant. A conversational interface to an AI model - you type, it replies. It answers and drafts, but on its own it does not take actions in other systems (that is an agent).
Context window. The maximum amount of text a model can consider at once, measured in tokens - the conversation, the documents you attach, and the instructions all share this budget. A bigger window means the model can “hold in mind” more at a time, at higher cost. When a chat gets very long, the earliest parts can fall out of the window and be effectively forgotten.
Data (training data vs. your data). Training data is the large body of text a model learned from before you ever used it. Your data is what you send at the moment of use (a prompt, an attached file). Knowing the difference is the heart of most AI privacy questions: what the model already learned versus what you are sharing right now, and where that goes.
Embedding. A way of turning text (or an image) into a list of numbers that captures its meaning, so a computer can measure how similar two things are. Embeddings are what let AI “search by meaning” rather than by exact keywords - the engine behind retrieval and recommendations.
Fine-tuning. Further training a general model on a focused set of examples so it gets better at a specific style or task. It changes the model itself, unlike a prompt or retrieval, which only change what you give it at the moment of use.
Foundation model. A large, general-purpose model trained on broad data that can be adapted to many tasks - the “base” that products, fine-tunes, and assistants are built on top of.
Grounding. Tying an AI’s answer to a trusted source - your documents, a database, a live system - so it responds from real facts instead of guessing. Grounding (often via retrieval) is the main practical defense against hallucinations.
Guardrails. The rules and controls that keep an AI system inside safe, appropriate boundaries - what it may discuss, which actions it may take, what data it may touch. Guardrails can live in the model’s training, in the system prompt, or in the surrounding software.
Hallucination. When a model states something false or made-up with complete confidence. It is not lying - it is predicting plausible-sounding text, and sometimes plausible is wrong. The fix is not to trust harder but to verify: ground answers in sources and check anything that matters.
In-context learning. A model’s ability to pick up a new pattern from examples you put directly in the prompt, without any retraining. Show it two or three examples of the format you want and it will usually follow them.
Inference. The act of running a model to get an output - every time you send a prompt and get a response, that is one inference. (Distinct from training, which is how the model was built in the first place.)
Jailbreak. A deliberate attempt to trick a model into ignoring its safety rules or instructions through clever wording. Relevant to anyone deploying AI: your guardrails have to hold up against users who probe them.
Large language model (LLM). The kind of AI behind today’s chat assistants. Trained on enormous amounts of text, it works by predicting the next piece of text given everything so far. That single idea - very good next-word prediction - is what produces its fluent answers, and also why it can hallucinate.
Memory. Whether an AI tool carries information from one conversation to the next. Most raw models are stateless (each chat starts fresh); products add “memory” features that store facts about you to reuse later. Worth knowing for both convenience and privacy.
Model. The trained AI system itself - the file of learned patterns that turns your input into an output. “Which model are you using?” is really asking which trained system (and which provider and version) is answering.
Multimodal. A model that handles more than just text - for example images, audio, or documents as input or output. A multimodal assistant can “look at” a screenshot or a PDF, not only read typed words.
Multi-turn conversation. A back-and-forth exchange over several messages, where each reply depends on what came before. Most real AI use is multi-turn, which is why the context window matters.
Open vs. closed model. A closed (proprietary) model is accessed only through a provider’s service; an open-weight model can be downloaded and run yourself. The trade-off is control and privacy versus convenience and, often, capability.
Parameters. The internal values a model learned during training - loosely, its “knobs.” Counts in the billions are common, and more parameters can mean more capability but higher cost. A rough size indicator, not a quality guarantee.
Prompt. What you send the model - your question, instructions, and any context. Clear prompts get better answers, which is why prompt-writing is a real skill.
Prompt engineering. The practice of writing prompts deliberately - giving context, examples, and a clear task - to get reliable results. Less arcane than it sounds: mostly it is being specific about what you want.
Prompt injection. An attack where hidden or malicious instructions (in a web page or a document the AI reads) hijack the model into doing something it should not. A key security concern once an AI can read outside content or take actions.
Provider. The company that hosts and serves a model (for example the makers of the major assistants). The same product may let you choose among several providers’ models.
RAG (retrieval-augmented generation). A technique that first retrieves relevant passages from your own documents, then feeds them to the model so its answer is grounded in that material. It is how an assistant can answer about your business without being retrained on it. RAG improves grounding but does not guarantee correctness - retrieval can be incomplete or stale, and the model can still misread what it is given, so verification still matters.
Reasoning model. A model tuned to “think” through a problem in steps before answering, trading a little speed for better performance on hard, multi-step tasks like math, coding, or planning.
Red teaming. Deliberately stress-testing an AI system to find where it fails or can be misused, before real users (or attackers) do. Standard practice for responsible deployment.
RLHF (reinforcement learning from human feedback). A training step where humans rate model responses and the model learns to prefer the better ones. A big part of why modern assistants feel helpful and polite rather than raw and erratic.
Safety training. The training that teaches a model to refuse harmful requests and stay within bounds. It reduces bad outputs but never eliminates them - which is why guardrails and human review still matter.
Shadow AI. Employees using unapproved AI tools (often free consumer apps) for work, outside any policy or oversight - a common data-governance risk that a managed AI offering is meant to solve.
Small language model (SLM). A compact model that is cheaper and faster than a large one, capable enough for many focused tasks and easier to run privately. Right-sizing the model to the job is a real cost lever.
Structured output. Getting the model to answer in a fixed, machine-readable format (such as JSON) instead of free prose, so its answer can feed directly into another system. Essential for automation.
System prompt. The hidden, standing instruction that sets an assistant’s role, rules, and tone for every conversation - set by whoever configures the tool, not the end user. It is a major lever for shaping and constraining behavior.
Temperature. A setting that controls how varied or predictable a model’s output is. Low temperature gives focused, consistent answers (good for facts); higher temperature gives more varied, creative ones.
Token. The unit AI models read and write in - roughly a word-piece (about four characters of English). Prompts, answers, and the context window are all measured in tokens, and most AI pricing is per token. “It costs by the token” is a useful thing to know when budgeting AI use.
Training data. The body of text (and other content) a model learned from before you used it. It shapes what the model knows and its blind spots, and it is separate from anything you send at the moment of use.
Vector database. A store built to hold embeddings and find the most similar ones fast - the retrieval half of RAG. It is what lets an assistant pull the right passage from thousands of your documents in an instant.
Zero-shot / few-shot. Zero-shot means asking a model to do a task with no examples; few-shot means giving it a handful of examples first. Few-shot often lifts quality noticeably for formatting or niche tasks - a simple, free improvement.