The model’s training data has a cutoff and no access to your files. So when an assistant answers about today’s news or your own documents, it is not remembering. It went and looked.
Under the hood, the app searches the web or your files, pulls the most relevant passages into the window, and asks the model to answer from them. The model reads what was found and writes from that.
Similar is not the same as right
Search ranks by relevance, and relevance means resembles the question. The passage that best resembles your question can still be outdated, from the wrong document, or subtly beside the point. When retrieval grabs the wrong thing, the model faithfully summarizes the wrong thing.
The quality of what comes back also depends on how you ask. Vague questions pull up vague passages. If you say exactly what you want, like the pet clause in the lease including anything about deposits, the search sharpens before the model writes a single word.
Why citations matter
A grounded answer can show its sources, and the sources are the answer’s receipts. A good habit: click through on anything that matters. If a claim has no source and matters, treat it as the model’s guess until confirmed.
Reading citations is a skill, and it takes about thirty seconds to learn. Check that the source is the kind that would actually know, meaning the lease itself and not a blog post about leases. Check the date. And check that the quoted line really says what the answer claims it says. Those three checks catch almost everything that slips through.
Knowing, finding, remembering
Three different capabilities get blended in every answer. Knowledge is patterns baked in at training. Finding is search at answer time. Remembering is the product’s saved notes about you. The best products are honest about which one an answer came from.
When an answer mixes all three, you can pull them apart by asking one question: how do you know that? A well-built assistant can tell you which parts came from your files, which came from a search, and which are general knowledge. If it cannot tell you, assume general knowledge, and verify anything specific before you rely on it.