What to do
1
Create a category
Group the fields, for example
Call Details.2
Add one extraction per field
Name is the result key; set it to your current key to keep names stable. Question is the sentence describing that field, lifted from your prose. Answer type is pre-defined if you listed allowed values, free text otherwise. Set expected format for numbers, dates, booleans, emails, or patterns.
3
Test against a real past transcript
4
Remove the old extraction task from the agent
Keys are now stable. The old feature let the model choose field names, so a key could appear or disappear between calls. The name you configure is the key you get on every call. If a call never covers the question, you get a low-confidence result rather than a missing key.
Before and after
call_reason and callback_requested moved from free text to pre-defined. Optional, but recommended: "Wants to reschedule" was whatever the model wrote that day; "Reschedule" is guaranteed to be one of your values.
What changes in your code
Full detail on the result object is in What the payload looks like, and the complete list is in Breaking changes.
Other features
Custom Analytics
custom_extractionsAgent Extraction
agent_extractionCall Summary
summaryexecution_id for a payload that does not match this guide.
Next Steps
Using Extractions
Configure extractions in the dashboard
Dispositions API
Manage extractions programmatically
Webhooks
Receive extraction data in real-time

