Define extraction fields

Custom extraction fields tell post-call analysis exactly what to pull out of each call. Every field becomes a key in call_analysis.custom_analysis_data, typed the way you defined it, so you can filter, chart and act on it without parsing transcripts.

Add a field in the dashboard#

Field types#

The model is constrained by a JSON schema built from your fields, so a selector can only return one of its choices, a boolean is always a boolean and a number is always a number.

Field settings#

The names call_summary, call_successful, user_sentiment and in_voicemail are reserved for the presets. A custom field with one of those names is ignored.

Required and optional fields#

An optional field is allowed to be null. When the model returns null, the key is removed from custom_analysis_data, so absent keys mean "could not be determined". A required field without a condition must be filled with a value of its type.

Conditions#

Only extract when limits a field to calls where something happened, for example a refund reason only when the caller asks for a refund. The condition is added to the field's instructions as "Only fill this when: ... Otherwise return null." A field with a condition is always nullable, even when Required is ticked. The field shows an If badge in the list.

In mock mode (no key for the analysis model's provider), a condition counts as met when at least half of its significant words (4 letters or more, minus common words) appear in the transcript.

Writing good descriptions#

The model only sees the field's name, type, description, examples and condition, plus the transcript and the disconnection reason. Write descriptions that:

• Say what to extract and from whose words: "The date and time the caller agreed to, as spoken."
• Say what to do in edge cases: "If the caller booked more than one appointment, use the first."
• Match the type: a Boolean description reads as a yes or no question; a Number description gives the unit ("Number of people in the booking").

Examples#

[
  {
    "type": "boolean",
    "name": "appointment_booked",
    "description": "True if the agent booked or moved an appointment during the call.",
    "required": true
  },
  {
    "type": "enum",
    "name": "call_reason",
    "description": "The main reason the caller called.",
    "choices": ["New booking", "Reschedule", "Cancel", "Question", "Other"]
  },
  {
    "type": "string",
    "name": "appointment_time",
    "description": "The appointment date and time agreed on the call.",
    "examples": ["Tue 21 Oct, 4pm"]
  },
  {
    "type": "number",
    "name": "party_size",
    "description": "How many people the booking is for."
  },
  {
    "type": "string",
    "name": "refund_reason",
    "description": "Why the caller wants a refund, in one sentence.",
    "conditional_prompt": "The caller asks for a refund"
  }
]

Set fields through the API with PATCH /update-agent/:agent_id:

curl -X PATCH https://api.tried.com.au/update-agent/agent_3f9c0a1b2c3d4e5f6a7b8c9d \
  -H "Authorization: Bearer $TRIED_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "post_call_analysis_model": "gpt-5.6-terra",
    "post_call_analysis_data": [
      { "type": "boolean", "name": "appointment_booked", "description": "True if an appointment was booked.", "required": true }
    ],
    "analysis_successful_prompt": "Successful only when an appointment was booked or moved."
  }'

post_call_analysis_data replaces the whole list, so send every field you want to keep.

Agent fields reference#

The same field shape (AnalysisField) is also used by extract variable tools and nodes, and by info collection steps, during the call. Post-call extraction runs after the call on the whole transcript.

Related#