The Model and the Full App
Work through the explanation, apply it to the example, and complete the quick check before continuing.
The model is not the whole product
A language model can receive tokens and generate tokens.
The full app may add many other parts around it.
If we mix all these parts together, we may wrongly believe that every model can search, remember, calculate, check, or act.
Let us separate the parts.
Search and retrieval
A search or retrieval tool can find current or approved material.
The found material may be added to the model's context.
This can improve an answer. It does not guarantee that:
- The right source was found.
- The source is correct.
- The model used it properly.
- The citation supports the claim.
Calculation and code tools
A calculator can give an exact result for a clear calculation. A code tool can run code instead of only writing it.
The model still needs to choose the right tool, send the right input, and use the result correctly.
Saved memory
An app may save user preferences or earlier details for later use.
This is a product feature. It is not proof that the base model remembers every user.
Saved memory also raises questions about consent, access, correction, and deletion. We will return to these issues in Module 6.
Citations
An app may attach citations from found documents. This is stronger than a citation made only from generated text.
Even then, a person may need to check that the source is real, suitable, and correctly used.
Agents
An agent can work towards a goal through several steps.
It may:
- Choose a tool.
- Take an action.
- Check the result.
- Change the next step.
- Continue until it reaches a stop point.
For example, a travel agent may search allowed options, compare them, ask for approval, and then make a booking.
The model may help choose the next step. The tools perform searches or actions. The full app controls access, limits, logging, and approval.
More action means more possible harm
A wrong chat answer may mislead a person. A wrong agent action may also change a calendar, send a message, update a record, or spend money.
We should ask:
- What can the system read?
- What can it change?
- Which actions need approval?
- Can a person stop it?
- Are actions logged?
- Who is responsible?
Calling a system an agent is not a safety check.
Compare three systems
| System | What it can do |
|---|---|
| Base text model | Generate text from available tokens |
| Tool-enabled assistant | Generate text and call allowed tools for a request |
| Agent | Work towards a goal through tools, actions, results, and next steps |
These are simple descriptions. Real products may sit between them.
Quick check
A model gives a current weather answer. What must be true?
A. Every language model has live weather data. B. The full app may have supplied current data, but we need evidence of how. C. The model's training must have happened today. D. The answer is correct because it is current.
Check the answer
Answer: B. A current answer may come from a search or weather tool added by the app.
Remember
- A base model does not automatically search, calculate, remember, cite, or act.
- The full app may add those abilities.
- Tool access does not guarantee correct tool use.
- Agents can take steps and actions towards a goal.
- More access and action need stronger limits, approval, logs, and review.
You are now ready to explain and check one complete AI output.