Skip to main content

​
Description

Use LLM Data when you need to extract one value from the current conversation. It is best for a single name, email, number, date, appointment time, reference number, or category. Use LLM Custom instead when you need several fields from the same answer.

​
Manual Inputs

NameData TypeRequiredDescription
TypestringYesWhat kind of answer to look for and how to format it
ReasoningstringNoLLM reasoning effort (low, medium, high). Defaults to low
LabelstringYesShort, noun-like name of the value being extracted
DescriptionstringYesThe extraction rule — what counts, what doesn’t, and the format
OptionsjsonNoAllowed choices (required only for struct-enum)
DestinationstringNoVariable to write the extracted value into

​
Type

Tells the tool what kind of answer to look for and how to format it.
TypeUse forOutput
struct-textNames, emails, registrations, reference numbers, notes, summariesFree-form text
struct-numNumeric valuesA float. Normalizes “fifty thousand”, “50k”, “1.5 million”; ranges use the midpoint
struct-dateA date onlyYYYY-MM-DD. Relative dates resolve from the call date; ambiguous numeric dates use DD/MM/YYYY
struct-datetimeA date and timeFull ISO datetime with the team timezone offset. Caller must give both clearly
struct-enumAn answer that must match one of your optionsThe matched option value

​
Reasoning

How hard the model works to interpret the answer.
  • low — direct answers (“My email is jane@example.com”)
  • medium — the caller may imply the value or answer less directly
  • high — genuinely subtle answers needing careful interpretation (can be slower)

​
Label

The short name of the value being extracted. Keep it noun-like; put rules in Description. 
Customer Email
Claim Number
Incident Date
Appointment Time
Customer Sentiment

​
Description

The main instruction. State exactly what counts, what does not, and how to format the result. This takes priority over the general type rules. 
Extract the customer's email address exactly as stated.
Extract the insurance claim number. Exclude policy numbers and phone numbers.
Extract the incident date in YYYY-MM-DD format. Resolve relative dates from the current call date.
Only pass when the caller gives both a date and a time.
For dates and datetimes, be explicit about whether partial answers are allowed. For reference numbers, say what similar values should be ignored.

​
Options

Used only for struct-enum. Provide the allowed choices the tool can return. Use stable, machine-friendly values: 
positive, neutral, negative
eligible, not_eligible, unknown
booked, declined, needs_follow_up
If you use labels or descriptions for the options, keep them short and distinct.

​
Destination

Optional variable to write the extracted value into. Leave blank if you only need the result boolean. 
{{contact.email}}
{{custom.claim_number}}
{{custom.appointment_datetime}}

​
Manual Outputs

NameData TypeDescription
resultbooltrue if a valid value was found and written (when a destination is set)

​
Conditions

​
Success (true)

A valid value was found and, if a destination was configured, written successfully.

​
Failure (false)

The value was missing, unclear, irrelevant, invalid for the selected type, or could not be written.

​
Example Usage

Extract an email: 
Type: struct-text
Reasoning: low
Label: Customer Email
Description: Extract the customer's email address exactly as stated.
Destination: {{contact.email}}
Extract an appointment time: 
Type: struct-datetime
Reasoning: medium
Label: Appointment Time
Description: Extract the appointment date and time only when the caller gives both.
Destination: {{custom.appointment_datetime}}
Extract sentiment: 
Type: struct-enum
Reasoning: medium
Label: Customer Sentiment
Description: Select the sentiment that best matches the customer's most recent answer.
Options: positive, neutral, negative
Destination: {{custom.customer_sentiment}}

​
Common Issues

  • Enum extraction fails — check that Options is filled in and the answer can reasonably match one option.
  • Date or datetime fails — check whether the caller actually gave enough information.
  • Returns false even though the value was extracted — check the Destination variable reference.
  • Inconsistent results — make Description more specific.

​
Next Steps