What the model knows, and what it needs

A confident answer is not a receipt

Imagine the assistant says, “Delivery is $2 and the restaurant is open.” Before trusting it, ask: where did those facts come from? If neither you nor a connected source supplied them, the response has not established they are current.

A large language model (LLM) learns patterns during training and generates responses from its available context. It can interpret “Thai would be nice” as a preference. That capability does not itself provide current restaurant availability.

Keep three things separate: instructions describe the task, user input describes the request, and external evidence supplies facts such as a current menu.

Check your understanding

Tokens are pieces; context is the working space

Models process text as tokens, which may be whole words, parts of words, or punctuation. A token is not always a word. The context window limits the information processed in a request; the exact accounting depends on the model.

Think of context as the material available on a desk. The current request, relevant prior messages, instructions, and tool results may all occupy space. A long conversation does not guarantee that every earlier message remains available.

When shortening a dinner conversation, preserve active requirements and corrections. Keep “vegetarian” and “$25 including fees”; remove irrelevant small talk first. Summaries can also make mistakes, so inspect them. See Anthropic’s context-window documentation for a provider example.

Check your understanding

Which information should survive when you shorten the conversation?

Tools provide a connection to the outside

An API is a defined way for software to request information or an action from other software. A tool makes a capability available to an assistant. The application must execute and check a tool request; naming a tool in a prompt does not connect it.

A future menu tool could supply prices and timestamps. For this module, label all restaurant examples as fictional. Do not claim that a working DoorDash integration exists.

This is a workflow: the application follows a defined sequence. We can add agent-directed choices when the problem needs them. This distinction follows Anthropic’s foundational workflow guidance.

Check your understanding

Match the information to its appropriate source.

Save your evidence audit

Record three claims, the source each needs, and a truthful fallback for a missing menu connection.

Key takeaways

• Model confidence does not establish that a fact was verified.
• Tokens are processing units, not a fixed synonym for words.
• Preserve current requirements when managing limited context.
• Tools need actual application connections and checked results.