Plain-English explainers

A spreadsheet has rows and columns. A database is like a spreadsheet with a third dimension: every table can link to every other table, so a name in one document connects automatically to every other place that name shows up, across a file with millions of entries, instead of you scrolling to find it by hand. Where a spreadsheet slows down past a few thousand rows, a database is built to hold millions and still answer a question in under a second.

PostgreSQL is the specific, free, open-source database program that stores and searches your file. It is the same category of software banks, hospitals, and government agencies run behind the scenes, not a consumer app, and it is what every build on this site actually runs on. Nothing about it is proprietary to me or locked to my equipment; your firm owns the copy of the software and the data once the build is delivered, and can move it to any machine.

OCR, optical character recognition, is software that reads the text in a scanned page image and turns it into words a computer can search. A scanned PDF looks like text but is really a picture of text, the same as a photograph of a page; OCR is what makes that picture searchable by adding a layer of real, selectable, searchable text behind the image, without changing what the page looks like on screen or in print.

A keyword search finds the exact words you typed. A meaning-based (vector) search finds pages that mean the same thing even when the words are different, the way a person who read the whole file would recognize that "fell and hit her head" and "sustained a head injury" are describing the same event. Both kinds of search run together, so you get exact matches and meaning matches at once.

A reranker is a second pass that looks at the search results a first pass already found and puts the most genuinely relevant ones at the top. Think of it as a research assistant who reads the pile a search engine handed back and hands you the five pages that actually answer your question, instead of a hundred that merely mention the words.

Bates numbering is the sequential stamp printed on each page of a production, so "page 4,217" always means the same physical page, no matter who is looking at it, which copy of the file they have, or how many years pass before someone cites it again. It is the standard way litigation keeps every party, and the court, pointing at the exact same page, and it is why the database ties every search result back to a Bates number instead of just a filename.

Instead of a chat window in a browser, the assistant runs in a plain text window on your computer, the same kind of window a technical person calls a terminal. You type a question in plain English, it searches your database behind the scenes, and it answers with the page it found the answer on. It is the same idea as a consumer chat AI, run against your own case file instead of the open internet, with citations instead of guesses, and nothing about it requires learning to write code.

A local language model is an AI model that runs on hardware sitting in your building instead of on a company's cloud server somewhere else. The tradeoff is that a local model is usually smaller and somewhat less capable than the largest cloud models, in exchange for a real guarantee: nothing in your file, and nothing in the question you ask about it, ever has to leave your office to get an answer.

A general cloud AI tool was not built on your file; it does not know your case, and when you paste case material into it, that material may leave your control and become part of a system not built for legal work. This build is the opposite: a database made only from your file, searched only for you, with every answer traceable back to a specific page.