Submit feedback on tasks

This guide shows you how to submit, list, and delete feedback for tasks and messages. Feedback is useful for building thumbs-up/down UIs, case review workflows, and automated evaluation pipelines.

Prerequisites

An existing task (see Send a message)
An access token or client credentials
A context ID and task ID

Submit feedback

# Replace these with your values
CONTEXT_ID="<your-context-id>"
TASK_ID="<your-task-id>"
ENVIRONMENT="<eu-or-us>"
TENANT="<your-tenant-name>"
TOKEN="<your-access-token>"

curl -X POST "https://api.${ENVIRONMENT}.corti.app/v2/agentic/contexts/${CONTEXT_ID}/tasks/${TASK_ID}/feedback" \
  -H "Authorization: Bearer ${TOKEN}" \
  -H "Tenant-Name: ${TENANT}" \
  -H "Content-Type: application/json" \
  -d '{
    "rating": {
      "scale": "binary",
      "value": 1
    },
    "labels": ["correct", "helpful"],
    "reason": "The response accurately identified the ICD-10 code."
  }'

Rating scales

Scale	Values	Description
`binary`	0 (negative) or 1 (positive)	Available now
`likert5`	1-5	Planned (not yet available)
`continuous01`	0.0-1.0	Planned (not yet available)

Labels

Labels provide structured observations about the result. Positive and negative labels can be combined to represent mixed feedback: Positive labels: correct, complete, helpful, wellPresented, efficient Negative labels: incorrect, missingInformation, irrelevant, misunderstoodRequest, unsupportedClaim, unsafeOrInappropriate, poorlyPresented, tooVerbose Neutral: other (requires a reason)

A maximum of five labels may be submitted per feedback entry. Duplicate labels are rejected. The other label requires a reason field.

Targeting a specific message

By default, feedback applies to the task as a whole. To target a specific message in the task’s history, include the target field with the message ID:

curl -X POST "https://api.${ENVIRONMENT}.corti.app/v2/agentic/contexts/${CONTEXT_ID}/tasks/${TASK_ID}/feedback" \
  -H "Authorization: Bearer ${TOKEN}" \
  -H "Tenant-Name: ${TENANT}" \
  -H "Content-Type: application/json" \
  -d '{
    "rating": {"scale": "binary", "value": 0},
    "labels": ["incorrect"],
    "reason": "The ICD-10 code was wrong for this specific response.",
    "target": {"messageId": "msg.0192f4c8-5f8d-7e61-924d-6fb09b5ead73"}
  }'

Metadata

Include provenance and correlation information using the metadata field:

{
  "metadata": {
    "collectionMethod": "thumbs",
    "clientReference": "session-abc-123",
    "actor": {"externalId": "user-pseudonymous-id"}
  }
}

Field	Description
`collectionMethod`	How the feedback was collected (e.g. `thumbs`, `survey`)
`clientReference`	Customer-defined reference for correlation
`actor.externalId`	Pseudonymous identifier for the submitter

The actor.externalId must not contain names, emails, national identifiers, or medical record numbers. Use a pseudonymous identifier.

List feedback

curl "https://api.${ENVIRONMENT}.corti.app/v2/agentic/contexts/${CONTEXT_ID}/tasks/${TASK_ID}/feedback" \
  -H "Authorization: Bearer ${TOKEN}" \
  -H "Tenant-Name: ${TENANT}"

The response includes all feedback entries for the task, scoped to the authenticated user via row-level security, newest first. The response is wrapped in a feedbacks array:

{
  "feedbacks": [
    {
      "id": "fb.0192f4c8-...",
      "taskId": "task.0192f4c8-...",
      "rating": {"scale": "binary", "value": 1},
      "normalizedScore": 1.0,
      "labels": ["correct", "helpful"],
      "reason": "The response accurately identified the ICD-10 code.",
      "createdAt": "2026-05-19T12:00:00Z"
    }
  ]
}

Feedback is scoped to the authenticated user. The response contains only the caller’s feedback, not all feedback for the task.

Delete feedback

Delete all feedback you submitted for a task. The DELETE endpoint targets the task as a whole — there is no per-entry feedback ID in the path:

curl -X DELETE "https://api.${ENVIRONMENT}.corti.app/v2/agentic/contexts/${CONTEXT_ID}/tasks/${TASK_ID}/feedback" \
  -H "Authorization: Bearer ${TOKEN}" \
  -H "Tenant-Name: ${TENANT}"

DELETE removes all feedback you submitted for the task, not a single entry. The operation is idempotent: deleting when there is no feedback returns 204.

Use cases

Thumbs UI: Collect binary feedback after each task response
Case review: Use labels like correct, incorrect, missingInformation for structured review
Automated evaluation: Submit feedback programmatically with metadata for tracking

Next steps

Learn how to work with contexts to find task IDs
Read about the task lifecycle for task states
Learn how to export traces for debugging

Quickstart

Concepts

Guides

Reference

Registry Connectors

Agent SDK

v1 Documentation (archived)

Prerequisites

Submit feedback

Rating scales

Labels

Targeting a specific message

Metadata

List feedback

Delete feedback

Use cases

Next steps

​Prerequisites

​Submit feedback

​Rating scales

​Labels

​Targeting a specific message

​Metadata

​List feedback

​Delete feedback

​Use cases

​Next steps

Prerequisites

Submit feedback

Rating scales

Labels

Targeting a specific message

Metadata

List feedback

Delete feedback

Use cases

Next steps