Chapter C06 · Interactive Course
What the app can actually see
Inspect what a particular interaction really has available.
Core Learning Concept
Pack the information needed for the task, then check how the app can access it. Distinguish text in the current request, a file actually attached, saved memory, and information retrieved through a tool. If a fact is missing, provide or retrieve it instead of assuming the model already has it.
What You’ll Learn in This Chapter
-
The app and the model are different
Distinguish a product's interface and features from the model inside it.
The model produces language predictions. The app is the surrounding product: its interface, file handling, memory, and tools. Two apps using the same model can offer completely different features depending on what tools are wired up.
-
Context is the material available now
Identify what information was actually supplied for the current response.
Context is the information available to a model for the current response. It may include instructions, recent messages, selected file content, and tool results. A fact that exists on your device is not necessarily included. Trace what was actually supplied before expecting the model to use it.
-
Long conversations can lose important details
Recognize that context capacity does not guarantee reliable use of every detail.
A context window limits how much information a model can receive at once. In a long conversation, an app may omit or summarize earlier material, and a model may fail to use a detail even when present. Restate the important requirement near the task and check the response against it.
-
Memory is a product feature
Distinguish current context from an optional saved preference.
Memory is a product feature for carrying selected information between interactions. It differs from the context supplied for one response and from training the model. Saving a preference does not mean every past conversation will be available. Check what the app says it remembers and correct stale details.
-
Files and pictures need actual access
Check whether the required material was provided and processed.
Mentioning a file is different from giving access to it. A model cannot inspect a file just because you named it. Attach the document and verify the app successfully read its contents before relying on its answers.
-
Internet access is not automatic
Look for actual retrieval capability and results when current facts matter.
A language model does not automatically browse the internet for each answer. A product can add search or browsing tools and decide when to use them. Look for evidence of an actual lookup and inspect the source. A confident answer about today’s events is not evidence of current access.
-
Sources can support an answer
Explain how retrieved material can inform generation without guaranteeing accuracy.
A source supports a claim when its actual contents justify that specific statement. A link can be real yet irrelevant or outdated. Open the source, find the relevant passage, and compare the details. If the source is narrower than the answer, narrow the answer instead of stretching the evidence.
-
Pack the right information
Choose an appropriate excerpt, distinguish memory from context, and verify whether an answer actually used a current source.
-
Explain an AI mistake
Inspect an answer containing an invented detail and an outdated fact, explain what information was available, and say what evidence is still needed.