The Smallest Extractor

one item, three fields

Part of: Meeting Notes to Action Items

A filter that finds candidate lines is a start, but "Alex will send the Q3 report by Friday" is still a sentence, not data. The smallest useful version of this product takes one commitment sentence and splits it into three named fields: owner , task , due . That's the atom everything else is built from. The extraction prompt We ask the model to do the splitting. The trick is telling it the exact output shape. Vague instructions ("pull out the details") give you a paragraph back; a precise format gives you something parseable. The simplest parseable format is a single delimited line: Why a delimiter first You'll upgrade to JSON in lesson 4, but a pipe-delimited line is the perfect training wheel. It's dead simple to produce and dead simple to parse: line.split(" "). Starting here keeps the two hard parts separate: first get the model to return a consistent shape , then worry about a richer shape. Learn one thing at a time. Parsing the reply Whatever format you ask for, the model's reply is just a string until your code gives it structure. Parsing the delimited line means splitting on the pipe and stripping the whitespace the model likes to add around separators: Now "Alex Send the Q

Challenge: Format the Extracted Items