"Grounded in your sources" is on the marketing page of almost every AI research tool now. Some of them mean it. Some run a light search, don't find much, and write a confident answer anyway, with citations attached that don't quite say what the answer claims.
Here is how to tell the difference in a few minutes.
Ask something your sources don't cover
Add your documents, then ask a question you know they don't address. A tool that's actually reading them will say so: no match, not covered, the sources don't discuss this. A tool that's guessing produces a fluent answer regardless. This one test filters out most of the weak ones.
Check that the citation supports the claim
Don't just check that a citation exists. Open it and read the passage. The failure mode isn't missing citations, it's a citation that points to a real page which doesn't actually back up the sentence it's attached to.
Ask for the exact wording
Request a direct quote, not a paraphrase. "Give me the sentence where the author says this." A tool built for retrieval can return the passage verbatim. A tool built to summarize keeps rewording it, which is a sign it's working from a compressed version rather than the text.
Ask the same question twice
Run the same query a few minutes apart. If the shape of the answer swings around a lot, the tool is leaning on the model's general reasoning, which varies between runs. If you're trying to work carefully, you want the answer to be stable across asks.
Why this matters
The objection people raise about chatting with your notes is fair: there's little value in an answer you have to fact-check against the source every time, because at that point you could have read the source yourself. The tools worth using close that gap. Every answer points to the card or passage behind it, one click to check, and the tool tells you when it doesn't know.
That is the design Graphia is built on. The AI reads everything on your canvas, answers from your material with the source attached, and says so when your sources don't cover the question. The point isn't to trust it blindly. It's that checking takes a click, not a re-read.