Ratings answer "how did this call go?". Extraction answers "what did they say?" — pulling defined fields out of free speech so you can filter, sort and export rather than read.
They are different jobs, and conflating them is the most common design mistake. A notice period is not a judgement; it is a fact. Scoring it out of five throws away the number you actually needed.
Defining fields that work
Extract facts, score judgements. Notice period, current city, expected compensation, willingness to relocate — these are extractions. Communication quality and technical depth are ratings. Keep the two lists separate and both get better.
Say the unit. "Notice period" invites "two months", "60 days" and "till end of June" — three formats you cannot sort. "Notice period in days, as a number" gets you 60. The instruction does the normalising so your spreadsheet does not have to.
Allow "not stated". This is the one that matters most. Without an explicit empty value, a model will infer something plausible from context, and an inferred salary expectation is worse than a blank — because a blank is honest and an inference looks like data. Say what to do when the topic never came up.
Constrain where you can. "Interest level: high, medium or low" beats "interest level" by a wide margin. Three values you can filter beat infinite prose every time.
What real speech does to your fields
People answer sideways. Asked about notice period they say "I can join after Diwali". Asked about salary they say "currently 12, looking for a decent hike".
Good extraction handles the first case by leaving it as stated rather than inventing a date. The second is a genuine judgement call you should decide in advance: does "a decent hike" become empty, or a note? Either is fine. Silently becoming 15 is not.
Using the output
The point is the filter. Once notice period is a number, "show me everyone under 30 days who is open to Hyderabad" is one operation instead of an afternoon of reading. That is the entire return on defining fields carefully.
Export to Google Sheets when the person who needs it does not use the dashboard — which, for hiring managers, is most of the time.
The check
Take ten calls you have read. Compare your extraction against what you would have written yourself. Disagreements are almost always a vague field definition rather than a model failure, and they are cheap to fix before the batch, expensive after.