There is a particular kind of lost. You know the document exists. You know roughly when it arrived and roughly what it concerned. You cannot remember what it was called, who sent it, or where it was filed — and every search tool you own wants exactly those three things.
So you search for a word you think might be in it, get four hundred results, and give up. Or you find it forty minutes later in a folder you would have sworn you had checked.
Why search fails here
Search matches words. You do not remember words. You remember circumstances.
“The letter about the boundary, from the solicitor, before we did the extension.” Nothing in that sentence is necessarily in the document. The word “boundary” may not appear; it might say “demarcation”. “Before the extension” is a date you would have to look up. “The solicitor” is a firm whose name you have half-forgotten.
Every one of those is trivially answerable by something that has read your archive, and impossible for something matching strings.
What the flow does
You describe the document the way you actually remember it, in a sentence, and get back a small number of candidates with a reason attached to each.
The reason matters more than the ranking. “This one, because it is from the firm you used in 2021, it mentions the boundary line, and it arrived two months before the planning application” lets you confirm or dismiss it in a second. A list of filenames does not.
The four descriptions that work
By relationship. “The thing the surveyor sent after the second visit.”
By consequence. “The letter that made us change the plans.” Often the only thing you remember, and completely invisible to search.
By position in a sequence. “The version before the one we signed.” Requires knowing which one you signed, which requires having read them.
By absence. “Did we ever get the certificate?” The most valuable question in this whole category, because the answer is sometimes no — and nobody ever searches for something they do not know is missing.
A worked example
“Find the insurance document that mentions the outbuilding. I think it was a renewal, maybe two years ago, and I remember it changed something about cover.”
A good answer is not a file. It is: the 2024 renewal schedule lists the outbuilding for the first time, and cover moved from the general limit to a separate one. The 2023 schedule does not mention it. There is also a broker email from the same month explaining the change.
You asked for one document. What you needed was the change and the reason for it, and only something that has read both years can see that.
What it should always do
Show you the source. Any claim about what a document says should come with the document. If it cannot show you, treat the answer as a lead rather than a fact.
Say when it is unsure. Three plausible candidates is a useful answer. One confident wrong one is worse than nothing, because you will act on it.
Tell you when the answer is no. “There is no certificate in what I have” is a real finding, and the one most likely to save you money.
The honest limit
It can only find what it has. If the document was never scanned, arrived as a photograph of a page at an angle, or lives in an account nobody connected, it does not exist as far as any assistant is concerned.
Which is the unromantic truth underneath all of this: the quality of the answer is set by the state of the archive, long before anyone asks a question.
What makes an archive usable →
Local AI. Private data. Local training.
$8,995 the first year — everything included. $4,995 each year after. The first year costs more because it contains the Jetson Orin placed and configured, your archive loaded and the first training run; every year after is the service running.
Or add the Archive Assessment first — $1,950, credited in full
You order and pay at checkout; we write back within days — a decline returns every dollar, and until our letter confirms the year you may withdraw in writing. From that letter the year is final, and it arrives on the date the letter names. Prefer to write to us first?