AI vocabulary has a way of showing up in product pages and pitch decks fully formed, as if everyone already knows what it means. Most of the time, nobody stopped to explain it. Here's the plain-English version of the terms that come up most.
Agentic — describes an AI system that doesn't just answer a question but takes multi-step action to accomplish something: checking a database, calling an API, deciding what to do next based on what it finds. A chatbot that answers from your docs isn't agentic; one that looks up an order, then decides to issue a refund, is.
Multimodal — a model that can work with more than just text: images, audio, sometimes video, alongside written language. A multimodal support tool could, in principle, read a screenshot a customer pastes in and respond to what's in it, not just what they typed.
Fine-tuning — additional training performed on top of an already-trained model, using a smaller, more specific dataset, to shift its behavior toward a particular task or tone. Most support chatbots don't need it — grounding in retrieved documents solves the "doesn't know my content" problem more directly and without retraining anything.
Zero-shot / few-shot — how a model is asked to perform a task with no examples ("zero-shot") versus a handful of examples included in the prompt to demonstrate the pattern ("few-shot"). Neither is about training the model — both happen entirely within a single request.
Inference — the technical term for "running the model to produce an output," as opposed to training it. Every chatbot reply is one inference call. It's also the stage most of a response's latency and cost comes from.
Hallucination — a model stating something false with the same confidence as something true, because nothing forced it to check. See grounded vs. hallucinated answers for the full explanation and how grounding prevents it.
Grounding — anchoring a model's answer to specific retrieved source material instead of letting it answer from general training alone. The core idea behind RAG.
Context window — the maximum amount of text a model can consider at once, conversation history and retrieved passages included. Covered in full in tokens and context windows, explained without the jargon.
Temperature — the setting that controls how much randomness goes into picking each word. Lower is more predictable, not automatically more accurate. Full explanation in what temperature actually controls.
Guardrails — the rules and checks layered around a model to keep it from doing things it technically could but shouldn't: answering outside its intended scope, leaking instructions it was told to keep private, or acting on a request it should refuse. Guardrails are engineering, not something the model does on its own.
Latency — how long it takes to get a response back. For a chatbot, this is the whole chain from question to streamed answer — see 300 milliseconds: everything that happens before a chatbot answers for what's actually happening in that time.
Related reading
- What is RAG, explained for non-engineers — the idea behind "grounding," in full.
- Grounded vs. hallucinated answers — the difference the "hallucination" entry is describing.
- 300 milliseconds: everything that happens before a chatbot answers — where "inference" and "latency" actually happen.