Ask Your File

Every lawyer, underneath the title, is a professional reader. The law buries you in paper. A single injury case can carry an eight-hundred-page medical record. A deposition runs for hours and lands on your desk as a transcript thick as a phone book. Contracts, discovery, exhibits, the file only grows. And somewhere in all of it are the three sentences that decide everything, waiting for you to find them by reading every line that isn't them.

You have dreamed, on some late night with a highlighter going dry in your hand, of simply asking the document. Where does this witness contradict himself? What did the doctor say about the second surgery? Is there a single line in this contract about who pays if the deal falls apart? Just ask, and have the document answer.

That is the tool we build in this chapter, and it is the first to put your private model to real work.

Why this one has to stay home

Notice what kind of documents these are. A medical record. A deposition. A client's contract. This is not the public, harmless level-one work from Chapter 4. This is level two and level three, the heart of it, packed with the most private facts your client owns.

Which means this tool is the whole reason you built the vault. You are about to feed real client material to an AI, and there is exactly one safe way to do that: on a model running on your own machine, with the document never leaving the building. No cloud. No vendor. No terms to trust. Just your file and your machine, in the same room, the way it should be. If you ever wondered what the private model was for, this is the answer. This is the day it earns its keep.

The danger, named plainly

Before we build, understand the one way this tool can hurt you, because it is the same way that has gotten real lawyers sanctioned.

When you ask an AI about a document, it can do something that looks helpful and is quietly dangerous. It can give you an answer that sounds exactly right, written with total confidence, about something the document never said. It can tell you the witness admitted fault on page two hundred when the witness said no such thing. It is not lying. It is guessing, and sometimes the guess is wrong, and it will never once warn you that it guessed.

The answer is a lead. The document is the evidence. You are still the lawyer. A good tool makes this easy by showing you its sources. A good habit makes you check them anyway.

Building it, which is mostly clicking

The happy news: you do not have to build much for this one, because GPT4All, the app you set up in Chapter 6, has a feature made for exactly this, and it shows you its sources right out of the box.

The feature is called LocalDocs, and it does one clever thing: it reads a folder of your documents, quietly, on your own machine, and builds a private index so the model can find the right passages when you ask. Your files never leave your computer. Setting it up looks like this:

1. Open GPT4All and find the LocalDocs section.
2. Click "Add Collection." Name it something like "Smith Medical Records."
3. Point it at the folder on your computer where those documents live.
4. Click create, and let it work for a few minutes while it reads everything in. When it shows a green "Ready," you are done.

Now open a chat, turn LocalDocs on, and ask your question in plain English. "What did Dr. Alvarez say about whether the second surgery was necessary?" The model answers, and underneath the answer is a button that says Sources. Click it. It shows the exact passages it pulled from, in the exact files. That Sources button is the most important button in this entire book, because it is the one that lets you do the checking the iron rule demands.

The discipline that keeps you safe

Now the part that matters more than any click.

Every single time the tool gives you an answer you mean to rely on, open the Sources and read the passage yourself. Let me be straight about why, because the reason is not theoretical. These document tools are useful but imperfect. On a clean, well-organized file they shine. On a messy scan, a bad copy, a document full of tables, they can stumble, and when they stumble they sometimes point you to a source that does not say what they claimed. The citation looks real. The passage, when you read it, is not what you were told. Trust the answer and skip the reading, and you walk into court with a fact that was never there.

So you read the source. Every time. It takes ten seconds, and it is the difference between a tool that saves you hours and a tool that ends your week in front of a judge explaining yourself. The AI finds the needle in the haystack. You are the one who looks at the needle and confirms it is a needle.

A note on cleaning up

Because this runs on your own machine, you do not have to strip the names and details out of these documents to be safe. The file never leaves the building, so there is no one out there to hide them from. That is the gift of working on your own model: the privacy is built into where the work happens, not into how carefully you scrubbed the file first.

One housekeeping note while we are here. Start with one clean, well-organized document, not your entire case file at once, so you learn how the tool behaves before you lean on it.

What you just learned

You did something today that a lot of lawyers are afraid to do: you handed real, privileged client material to an AI, and you did it safely, because it never left your machine and because you checked every answer against its source. That is the whole game with documents. Keep it home, and never trust what you cannot trace.

You also have, sitting on your computer right now, a private research assistant that reads faster than any associate and forgets nothing. Point it at a record and ask. Just remember to look at the needle.

In the next chapter the model stops only reading and starts writing. You will build the First-Draft Engine, a tool that takes the letter you have written a hundred times and writes the first version for you, and leaves you to do the one thing it cannot: be the judgment that decides what goes out the door.