If you start reading about how to make an assistant know your own material, you will quickly meet two pieces of jargon: retrieval and fine-tuning. They are usually presented as rivals, and the argument between them is usually conducted by people who enjoy the argument.
You do not need the jargon. You need to understand two different things a system can do with your papers, because they fail in different ways and a good arrangement uses both.
The first way: looking it up
The first method keeps your material as material. When you ask a question, the system searches your documents, finds the passages that seem relevant, and reads them before answering.
This is closer to a very fast assistant with a very good filing system than to anything mysterious. The model itself has not changed. It has simply been handed the right pages at the right moment.
What it is good at. Facts that must be exactly right. A figure, a date, a clause. Because the system is reading the actual document, it can show you the source, and you can check it. It also updates instantly: add a document this morning and it can be used this afternoon.
Where it struggles. Questions that are not about one passage. “Are we drifting over budget on this project?” is not answered by finding the right page, because no page says it. The answer lives across hundreds of documents, and looking things up is poorly suited to that shape of question.
The second way: learning it
The second method changes the model itself. Your material is used to adjust how it responds, so that the way your world works becomes part of how it thinks rather than something it consults.
The result is harder to describe and easy to recognise. An assistant that has learned your material knows your vocabulary without being told, knows which supplier you mean when you use a nickname, knows the shape of your year, and writes in a way that sounds like your practice rather than like a press release.
What it is good at. Pattern, tone, judgement and the unwritten. The things everyone in your household or firm knows and nobody has ever written down.
Where it struggles. Exact recall. A model that has learned from a thousand invoices has an excellent sense of what your invoices are like and should not be trusted to recite the total of invoice 407 from memory. Ask it to, and it may produce a plausible number. That is the failure worth understanding, because it is confident and wrong rather than blank.
Why the argument is a false one
Set out like that, the answer is obvious: you want both, doing what each is good at.
Learning gives the assistant a sense of your world — the vocabulary, the habits, the judgement about what matters. Looking up gives it the facts, with the source attached, so that any specific claim can be checked against the document it came from.
A system built on only the first is a stranger with a good search box. A system built on only the second is a colleague with a confident memory and no filing cabinet, which is a genuinely dangerous combination for anything involving numbers.
What this means for you
Three practical consequences, if you are evaluating anything in this category.
Ask what happens to a number. If you ask for a figure, does the answer come with the document it came from? If it cannot show you, treat every number it gives you as a draft.
Ask how new material arrives. Material you add should be usable quickly. If everything must wait for the next training run, the assistant is always slightly out of date, and out of date is worse than it sounds when the subject is your own affairs.
Ask where each part happens. Both methods require something to read all of your material. That is the privacy question, and it is settled by where the computer is, not by which method is used.
The part that is the same either way
Both methods depend entirely on the state of your material. Neither one rescues an archive full of contradictions, duplicates and unreadable scans — looking up will find the wrong version, and learning will learn the wrong version.
Which is why the first useful step is not choosing a method. It is finding out what you actually have.
What can and cannot be trained on →
Get a written scope of your own material →
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?