Skip to main content
Takeaways turn finished conversations into validated JSON. Define a takeaway once — a name, a JSON Schema, optionally a custom prompt and model — attach it to any number of agents, and after each call an LLM fills the schema from the transcript. Results arrive inside call.ended and the analysis API. Use them for CSAT scores, lead qualification fields, booking details, follow-up flags — anything you’d otherwise parse out of transcripts by hand.

​
Create a takeaway

curl -X POST http://localhost:8090/v1/takeaways \
  -H "Authorization: Bearer tc_xxx" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "csat",
    "description": "Customer satisfaction, inferred from the conversation",
    "schema": {
      "type": "object",
      "properties": {
        "score": {"type": "integer", "minimum": 1, "maximum": 5},
        "reason": {"type": "string"}
      },
      "required": ["score"]
    }
  }'
The name keys the result in payloads, so it must be identifier-shaped (csat, lead_quality, booking-details) and is immutable after creation. The schema is validated at create time — a broken schema fails here, not after every call. Optional fields: prompt (extra extraction instructions) and model (per-takeaway LLM override — e.g. a cheap model for a boolean, a strong one for complex extraction).

​
Attach to agents

Reference takeaways by id in the agent’s analysis config: 
{
  "analysis": {
    "enabled": true,
    "takeaway_ids": ["<csat-takeaway-uuid>", "<lead-quality-uuid>"]
  }
}
Multiple agents can attach the same takeaway; editing it in one place updates extraction everywhere.

​
Results

Each attached takeaway is extracted in its own concurrent LLM call after the call ends (a few seconds, in the background). Results land keyed by name: 
// call.ended payload → analysis.takeaways
{
  "csat": {
    "result": {"score": 4, "reason": "Issue resolved quickly"},
    "valid": true,
    "model": "gpt-4o-mini",
    "duration_ms": 1240
  }
}
Also available via GET /v1/calls/{call_id}/analysis. Extracted JSON is validated against your schema (with one automatic retry on mismatch); a failed extraction is reported as {"valid": false, "error": "..."} rather than silently dropped — one bad takeaway never affects the others.

​
Manage

POST   /v1/takeaways          # Create
GET    /v1/takeaways          # List
GET    /v1/takeaways/{id}     # Get
PUT    /v1/takeaways/{id}     # Update (schema/description/prompt/model — not name)
DELETE /v1/takeaways/{id}     # Delete (409 while agents still attach it)
The older inline analysis.structured_extraction_schema keeps working (result at analysis.structured_data), but takeaways are the recommended path — reusable, multiple per agent, independently validated.