AI conversations get easier once you can distinguish the model from the product around it, and a generated answer from an action. Most of the jargon describes one of those pieces.
Imagine a bicycle shop wants help answering questions and arranging services. Here is what the familiar terms mean in that setting.
The model is one part of the assistant
A model is the trained system that processes information and generates a response. An assistant is the product built around it, including the interface, instructions, reference material, tools and permissions.
Two assistants using the same model can give different answers because one has the shop’s current service list and the other does not. Choosing a model is therefore only part of choosing a useful product.
Generative AI creates outputs such as text, images or audio. It can draft a service description; it does not acquire the authority to change your offer simply because its wording is persuasive.
Tokens, context and memory
Tokens are units a model processes. In text they may be words, parts of words or punctuation; the relationship varies. Product limits measured in tokens are not reliable word counts.
The context window is the capacity available for information during processing. A bigger window makes room for more material, without guaranteeing that every detail will be used correctly.
For the shop, context could include the customer’s message, a service list and earlier conversation. If a qualification is buried in a long document, it still needs checking.
Memory is a product feature that can carry selected information into later interactions. It is different from the context window. Remembering your preferred tone does not establish that the assistant remembers this week’s prices.
How an assistant looks up your business information
A model may know a lot about bicycle maintenance, but that does not mean it knows what this particular shop includes in a service. To answer questions about the business, the assistant needs access to its information.
Suppose a customer asks, “Does your standard service include replacement brake pads?” In our example, the shop’s servicing policy says that labour is included, while replacement parts are quoted separately.
Retrieval is the step that finds the relevant information—in this case, the paragraph about parts and labour. The system then gives that paragraph to the model alongside the customer’s question, so it can draft an answer explaining what is included and what would cost extra.
That combination of looking up information and using it to generate an answer is called retrieval-augmented generation, or RAG. The reference material could come from documents, a product catalogue or another connected source. AWS’s explanation of RAG describes how this supplies information without retraining the model.
For the shop, the benefit is that its policy can remain the maintained source of truth. When the policy changes, the information available to the retrieval system needs updating too. Otherwise, the assistant could keep answering from an old version.
You may also hear fine-tuning discussed as a way to adapt AI to a business. This means further training a model using examples to change how it responds. For instance, examples might help it follow a consistent format for service summaries. Retrieval supplies reference information for an answer; fine-tuning changes the model through training. They serve different purposes and can be used together.
Neither removes the need to check the result. The search could miss the paragraph about extra charges, or the model could misread it. A useful check is to compare the answer with the actual policy, including any conditions or exceptions.
Tools turn a conversation into a workflow
A tool gives the assistant a capability, such as searching a service catalogue or checking a calendar.
An agent uses a model to choose actions across steps. It might identify a question, consult the catalogue and prepare a proposed booking.
The important distinction is permission. Reading available appointments, reserving a slot and charging a deposit are separate actions. The word “agent” does not tell you which are allowed.
An API is a defined way for software systems to communicate. An assistant might check availability through an API rather than clicking through the booking screen. Either route still needs appropriate access.
Multimodal does not mean all-knowing
Multimodal systems work with different types of information, such as text, images and audio. A customer could show a bicycle photo and ask a question.
Seeing the image does not establish the condition of a hidden component or qualify the assistant to assess mechanical safety. The input is richer; the evidence still has limits.
A hallucination is incorrect or unsupported generated information. If an assistant invents a service inclusion, the polished sentence does not make it part of your offer.
You can understand most new AI features by asking three questions: what information can it use, what output can it produce, and what can it change? Ask those before worrying about the next new acronym.
Read next: build a first AI workflow.