S3E4 The Extractor

What this lesson is about

A PDF is a picture of a document. A photograph of a worksheet is a picture of a picture of a document. Neither can be searched, quoted, fed to another tool, or pulled into the harness you have been building all season. The Extractor's whole job is to turn any of those back into plain text you can actually use. You give it a file or a link; it gives you clean markdown. That is the entire job, and it does nothing else on purpose.

What you'll be able to do

• I can get the full content out of a PDF, a photograph, a slide deck or a spreadsheet as structured text.
• I can explain why a specialist tool that does one job well beats a general assistant that will attempt anything.
• I can judge an extraction by checking it against the original, rather than by whether it reads well.
• I can save what I extract into my own toolkit so a later session can use it.

A little background

This episode closes the season's loop. Season 3 has been about building context for an AI to work from — first a file describing you, then rules for an agent, then the retrieval step that searches those files. The Extractor is what fills them. It is the difference between a folder of documents you own and a folder of documents an AI can read.

Extraction is accuracy work rather than reasoning work, so the way to judge it is not whether the output reads nicely. It is whether the output matches the original. Pick something off the page you can see with your own eyes — a figure in a table, a heading halfway down, the last line — and find it in what came back. If those match, the extraction held.