An assistant that answers questions about your own files is genuinely useful. It is also, by default, a fluent guesser. Ask it something it doesn't know and it won't shrug; it invents a plausible number in the same confident voice it uses for the true ones. On a movie pick, that's fine. On the day your insurance lapses, a smooth guess is worse than a blank. At least the blank tells you to go look. So Findy runs on one rule.
No passage, no answer
Every reply it does give arrives with a Receipt: the file, the page, and the exact words it stood on. That's the highlighted line you saw above. Pinning an answer to your own words is called grounding, and it's the difference between a system that knows and one that guesses.
Four gates, and none of them trust the model
You just watched them run: the question has to be real; the search has to find something in your files; those matches have to carry real evidence; and the answer has to cite a document. Miss any one and the reply is dropped without a word.
The fourth gate is the one I care about. The answer model is Google's Gemini, run at temperature 0 so it doesn't wander, and it has to hand back two things: the answer, and the id of the file it used. That id gets checked against an allowlist, the only documents actually put in front of it. Cite a file it was never shown, or cite nothing, and the whole answer is thrown away. In the code, the rule is literal: an uncited answer does not exist.
Allowed to find. Never to invent.
To find your files fast, Findy keeps a short AI-written summary of each one and searches that too. It's good at locating a document by meaning. It's also blocked, by design, from being the source of an answer. So a reply can only ever come from words that are really in your file, never from the machine's own paraphrase of it.
Find the right file by what it's about.
Answer. Only your document's real words can.
Refusing is the expensive choice
A model will always produce text; that's what it's for. The months went into the machinery that lets Findy say nothing, and say it without ceremony. When it has no answer, the box just collapses. You don't get an apology, or the “I'm sorry, as an AI.” It stops, and moves on.
Getting it to answer was the easy afternoon. Getting it to shut up cost me the year.