"Grounded" is a marketing word right up until the moment you can check it. Every AI product on the market says it will not invent things. What separates them is whether they hand you the means to catch it when it does.
So rather than repeat the claim, here is the machinery: one question, from the moment you press Enter to the buttons sitting under the finished answer.
Watching it arrive
You type, you send, and the answer streams in word by word while a thinking indicator holds the space. Nothing unusual there. Two moments are worth noticing.
The first is the send button turning into a Stop Generating square. Stopping keeps whatever has been written so far instead of throwing the answer away, which is useful when the opening paragraph already told you what you needed.
The second is rarer and easy to miss. Sometimes the assistant pauses mid-flow and shows "Taking a moment to think this through…". It is working your question a second way behind the scenes. It needs nothing from you, and the answer that follows is usually the better one.
What you cannot do is pick a model. That choice belongs to the platform and the document owner, deliberately, because a reader should not have to know what a model is. What you can choose is length. Every answer carries a badge reading Chatty or Succinct, and tapping the badge redoes the answer in the other style, in place, instead of piling a second version underneath. Your choice wins even when the owner's custom prompt has asked for long, thorough answers. That precedence is the right way round: the person reading decides how much they want to read.
The citation is the product
As it writes, it cites. Small Ref chips appear inside the text, and the full list gathers at the end.
Those chips carry the whole argument. Select one and a source viewer opens onto a card labeled Source text, holding a verbatim excerpt from your document. Not a paraphrase, and not a summary the model wrote about its own reasoning. Above the excerpt sit the reference number and the Page it came from, plus the document name when you are chatting across several at once. Where one citation draws on more than one passage, arrows step between them.
Plenty of assistants will tell you they read your document. This one shows you the paragraph. A reader who doubts a sentence is two clicks from the page number, and a reader who doubts nothing never opens the viewer at all.
Two things to know. Owners can switch the reference display off, so an answer without chips is a setting rather than a failure. And these numbered citations are a different thing from the References tool beside the chat, which is a bibliography the owner writes by hand. If you want the original file instead of an excerpt, that is the Files tool. Both live in the tools that sit beside the chat.
The refusal is the feature
This is the behavior to test before you trust anything else. Ask a trained assistant something its document does not cover, and you get a heading that reads Not in this document, followed by a plain admission:
I couldn't find that in my training. I only answer based on what's in these materials. I won't guess beyond that.
No hedge, and no plausible-sounding paragraph you then have to go and verify. Most of the value in a grounded assistant is concentrated in that one response, because it is what lets you believe the other ones.
It is not a dead end, though. If the document belongs to an organization with others, the chat can search those too and offer an Elsewhere panel, "This may be answered in other documents", listing candidates with their category, year and page. Ask there opens that document and re-asks your question automatically. Inside a collection, the same panel is limited to that collection's documents, and only the ones you are allowed to open, so a helpful suggestion never quietly walks a reader out of the set they were given.
Answers are not only text
Images from the source document load into the answer and open full screen when tapped, with captions in a footer band underneath, laid out like a figure in a published paper. Where an owner has added a video gallery, a strong match can bring a playable card into the answer itself. Images and video covers both.
When a question could reasonably go two ways, the assistant answers what it can, then asks. "Which population would you like?", with tappable chips beneath it. Tap Adults and it reads the two-word reply in the context of what you were already discussing. The chips are shortcuts rather than a menu you are trapped inside.
What happens after you read it
Under each answer sits a row of small buttons: copy it, export it as a PDF by email, re-ask and replace it, or share the conversation as a public link. Then thumbs up and thumbs down.
The thumbs-down is worth more than it looks. It opens a short form, What went wrong?, with seven reasons you can combine, from Factually inaccurate to Missing references. Those reasons are not politeness. They arrive in the owner's conversation viewer as tags, which makes them the most direct signal available about which parts of a document are thin. See conversations and moderation.
On medical documents, a small hazard icon can appear alongside: a caution to verify medication details before prescribing, or a notice when a question reads like an adverse event report. That second flag sometimes arrives a few seconds after the answer settles.
One last thing. Where an owner has given their assistant a name and a face, answers arrive as "(Name) Says:". That is presentation, and only presentation. The paragraph underneath is still assembled from the document, and the chips inside it are still the thing to check.