Back to catalog
AI at workBeginner

AI for Documents and Data

Uploading files to a chat: what the model really sees, how to pull data out into a structure, and how to prove nothing was lost

AI for Documents and Data

About this course

A course for everyone who receives work as files: contracts and invoices, statements and reports, scans, price lists, exports from systems, forty applications in free form. The question is always the same — how to get from a pile of documents to rows you can work with, without retyping them by hand and without quietly losing half of them on the way.

We start with the part almost nobody is told: a file you attached does not reach the model the way it looks on your screen. A scan without recognition is a blank sheet, a merged cell becomes an empty one, a running head turns into a sentence in the middle of a paragraph. Then come the working techniques — extraction into named columns with a "not found" instead of a guess, parsing in batches with a row count, reconciling two documents field by field, reading a long document with quotes you can verify by searching the file, cleaning and categorising table data.

The honest part runs through every lesson. A model extracts and structures well, but it never guarantees completeness — and no wording makes it guarantee that. So each technique comes with its control: counters, an "unparsed" list, a mandatory quote column, a spot check against the original. No programming, no product names, no model versions.

Course curriculum

  1. What a file turns into on the way to the model

    Why an attached document does not reach the model the way it looks on your screen, how to tell in thirty seconds whether it was read at all, and what must never be uploaded.

    • 1.What the model actually sees in your file11 min
    • 2.A scan without recognition is a blank sheet12 min
    • 3.What must not be uploaded to someone else's service10 min

  2. Extraction into a structure you set

    Getting named fields out of documents without invented values, parsing a pile of text in batches while keeping count, and reading a long document with quotes you can verify by searching the file.

    • 1.Extraction by a given structure: fields, rules, "not found"14 min
    • 2.A pile of text into rows: batches and keeping count13 min
    • 3.A long document, a purpose, and quotes you can search for15 min

  3. Two documents, many files, and data in columns

    Comparing documents field by field instead of asking whether they match, keeping several files apart in one conversation, and preparing table data without letting the model do arithmetic.

    • 1.Reconciling two documents: where they disagree14 min
    • 2.Several files in one conversation12 min
    • 3.Table data: clean and categorise, do not compute13 min

  4. Limits, and proving that nothing was lost

    Where file work reliably breaks — size, layout, merged cells, running heads and footnotes — and the three checks that turn a plausible result into a result you can sign off.

    • 1.Where it breaks: size, layout and the things that look like text13 min
    • 2.Three ways to prove nothing was lost16 min