An AI application may retrieve documents and then use a model to compose a response from them. The retrieval step selects material; the generation step produces wording using the supplied context.
A retrieved source does not guarantee that every sentence accurately reflects it. Check whether the cited page supports the particular claim, and distinguish information from the source from an inference added in the answer.
Open important references and read the relevant passage in context. The combination can be useful, but each stage has its own possible gaps. Understanding the difference helps explain why an answer with links still needs a check.
Imagine a practical setting.
A relevant search result and a generated conclusion are separate objects. Keeping the source visible helps a reviewer examine the connection between them.
Before the next experiment.
A small trial should have a clear stopping point. Decide which uncertainty the tool can help explore and what observation would answer the next question.
- Distinguish retrieval from composition.
- Check support for each important claim.
- Identify inferences added beyond the source.
Follow a related question
Record the reason for the shortcut.
A shortcut with a later costRecord capture time and covered period.
A snapshot needs its dateKeep learning
Related background to continue exploring this subject.
Google: an introduction to language models NIST: AI risk management framework