Collecting and Publishing User Feedback

Solace Agent Mesh can collect end-user ratings on agent responses, store them, and expose them through a REST API for analytics and agent evaluation. Feedback collection is available on the Agent Mesh UI, Slack, and Teams entrypoints, and every entrypoint writes to one shared feedback store. This page covers the operator configuration. For the end-user experience of rating a response in the UI, see Giving Feedback.

Feedback moves through the following stages, each configured independently:

  • Collection stores each rating, its optional comment, and the end user's content-sharing consent in the entrypoint database. Enable it per entrypoint.

  • Publishing forwards Agent Mesh UI feedback to the event mesh so an external consumer can act on it.

  • Retention prunes stored feedback on a schedule so the database does not grow without bound.

Before You Start

Feedback collection requires a persistent session store, because the entrypoint saves each rating to its database. Configure the session service to use SQL persistence before you enable feedback. For how to configure session storage, see Session Storage.

All feedback collection requires the SQL session store. Without it, the Agent Mesh UI hides the rating buttons, and the entrypoint logs a warning when the UI requests its configuration, while the Slack and Teams entrypoints still show their buttons but discard submitted feedback. In every case, the end user sees no error.

Enable Feedback Collection

Feedback collection is enabled per entrypoint. Each entrypoint type has its own setting, and every entrypoint writes to the same feedback store. The Slack and Teams entrypoints are configured as declarative configuration resources; for the sam config workflow, see Managing Configuration as Code (Early Access). The Agent Mesh UI examples show individual configuration keys. When configuring the Web UI entrypoint, add these keys to the same configuration file or section where you set other Web UI entrypoint settings.

Agent Mesh UI

Set frontend_collect_feedback to true in the Web UI entrypoint configuration to show the Like and Dislike buttons on each agent response:

# entrypoint configuration
frontend_collect_feedback: true

As an administrator, you can also set collectFeedback through PUT /api/v1/platform/webuiSettings or the manifest's platform.webuiSettings block. An override set there takes precedence over frontend_collect_feedback in the entrypoint's own YAML. For more information, see Web UI Settings.

Slack

In the Slack entrypoint configuration, set feedback_enabled to true under spec.values to add thumbs-up and thumbs-down buttons after each response:

# entrypoints/slack-entrypoint.yaml
kind: entrypoint
name: slack-entrypoint
description: Slack entrypoint with user feedback enabled.
spec:
  type: slack
  values:
    slack_bot_token: ${SLACK_BOT_TOKEN}
    slack_app_token: ${SLACK_APP_TOKEN}
    feedback_enabled: true

For the complete Slack entrypoint setup, see Slack Entrypoints.

Teams

In the Teams entrypoint configuration, set feedback_enabled to true under spec.values to post a thumbs-up and thumbs-down card after each response:

# entrypoints/teams-entrypoint.yaml
kind: entrypoint
name: teams-entrypoint
description: Teams entrypoint with user feedback enabled.
spec:
  type: teams
  values:
    feedback_enabled: true

For the complete Microsoft Teams entrypoint setup, see Microsoft Teams Entrypoints.

Feedback Records

Each stored feedback record holds the rating (up or down), an optional comment, and a content-sharing flag. The flag records whether the end user consented to let an administrator view the conversation content alongside the feedback. This consent is off by default, and only the task owner can grant it.

Feedback comments are stored in the entrypoint database, are kept for the full retention window, and, for the Agent Mesh UI, are included in published feedback events. Comments and shared task content can contain personal or sensitive data, so weigh this risk before you enable collection or publishing.

Publish Feedback to the Event Mesh

Publishing applies to Agent Mesh UI feedback only. The Slack and Teams entrypoints store feedback and expose it through the API, but do not publish it to the event mesh.

Publishing is off by default and requires the SQL session store described in Before You Start, because only stored ratings are published. Turn it on in one of two places.

The entrypoint no longer honors feedback_publishing.topic. Feedback publishes to the fixed namespaced topic described in The Feedback Event, and a configured topic value is ignored with a startup warning. If you set this key in an earlier release, move your subscriber to the fixed topic.

An administrator sets it per instance through the Platform service, and the change takes effect without restarting the entrypoint. Send a PUT request to /api/v1/platform/webuiSettings:

{ "collectFeedback": true, "publishFeedback": true }

You can declare the same settings in a configuration manifest instead:

# manifest.yaml
platform:
  webuiSettings:
    collectFeedback: true
    publishFeedback: true

An operator sets the deployment-wide default in the Web UI entrypoint configuration. A platform override, when present, takes precedence over this value:

# entrypoint configuration
frontend_collect_feedback: true
feedback_publishing:
  enabled: true
Field Description
enabled When true, the entrypoint publishes a feedback event every time a rating is submitted. Defaults to false.

The Feedback Event

The entrypoint publishes every rating to one topic, where {namespace} is its configured namespace:

{namespace}/sam/v1/feedback

The topic is fixed. To route feedback into your own topic taxonomy, republish from this topic with a bridge on the event broker.

{
  "feedback": {
    "event": "created",
    "task_id": "3b1e5c7a-9d21-4f0a-bd11-6a2e4c8f9012",
    "session_id": "9c2a1f6e-4b83-47d2-a0c5-1e7d9b3f5a24",
    "feedback_type": "down",
    "feedback_text": "The summary missed the breaking changes.",
    "user_id": "user-42",
    "entrypoint_type": "webui",
    "subject_id": "OrchestratorAgent",
    "subject_type": "agent",
    "created_at": "2026-09-09T14:22:31.004000+00:00",
    "updated_at": "2026-09-09T14:22:31.004000+00:00"
  }
}

The event field is created the first time a user rates a response and updated if that user later changes their rating. Both events carry the same task_id and user_id, so a consumer that counts ratings must treat the pair as one rating and take the latest.

The published event always carries feedback_text and user_id, regardless of the content-sharing flag on the record. That flag controls whether an administrator can view the conversation content through the API. A subscriber to this topic receives comments that the API redacts for administrators without content access. Grant subscribe access on this topic only to the audience you trust with feedback content.

Agent Mesh does not consume this event itself. Subscribe to the topic from your own service to route feedback into a data warehouse, an alerting system, or an evaluation pipeline. To confirm events are flowing, watch the topic while you submit a rating in the Agent Mesh UI.

The event carries the user's comment and their user ID, and the event broker delivers it to every subscriber it allows on that topic. Review who can subscribe within the namespace before you enable publishing.

Set Feedback Retention

The same automatic sweep that prunes tasks and event history also prunes stored feedback, no matter which entrypoint collected it. Set feedback_retention_days under data_retention to control how long feedback is kept:

# entrypoint configuration
data_retention:
  enabled: true
  feedback_retention_days: 90
  cleanup_interval_hours: 24

The sweep deletes feedback older than feedback_retention_days on every run. The default retention is 90 days, and the default sweep interval is 24 hours. For more information about the retention sweep and the other data it prunes, see Managing Backups and Data Retention.

Retrieve Feedback Through the API

Agent Mesh exposes stored feedback at GET /api/v1/feedback. Feedback collected on any entrypoint that shares this database is available here, and each caller can retrieve only their own feedback:

curl -H "Authorization: Bearer ${SAM_AUTH_TOKEN}" \
  "https://myapp.example.com/api/v1/feedback?rating=down&pageSize=50"

This endpoint accepts these query parameters:

Parameter Description
startDate, endDate Restrict results to feedback created in a time range. Both are ISO 8601 timestamps.
taskId Return feedback for a single task. The caller must own the task.
sessionId Return feedback for a single session.
rating Filter by rating value, either up or down.
pageNumber The 1-based page to return.
pageSize The number of records per page, between 1 and 100.

The response wraps the records in a paginated envelope:

{
  "data": [
    {
      "id": "f47ac10b-58cc-4372-a567-0e02b2c3d479",
      "sessionId": "9c2a1f6e-4b83-47d2-a0c5-1e7d9b3f5a24",
      "taskId": "3b1e5c7a-9d21-4f0a-bd11-6a2e4c8f9012",
      "userId": "user-42",
      "rating": "down",
      "comment": "The summary missed the breaking changes.",
      "shareTaskContent": true,
      "createdAt": "2026-07-03T14:41:54.021000+00:00"
    }
  ],
  "meta": {
    "pagination": {
      "pageNumber": 1,
      "pageSize": 50,
      "count": 1,
      "totalPages": 1,
      "nextPage": null
    }
  }
}

The shareTaskContent field reflects the consent the end user gave when submitting the rating. When it is true, an administrator can review the associated task content alongside the feedback.

Troubleshooting

Rating Buttons Do Not Appear, or Clicking Them Records Nothing

Feedback collection is enabled, but the buttons or card are missing, or submitting feedback stores nothing.

  • Confirm the entrypoint's session service uses type: sql. Without a SQL store, the Agent Mesh UI hides the buttons, while the Slack and Teams entrypoints show their buttons but discard the feedback.

  • Confirm the enable setting for that entrypoint type: frontend_collect_feedback for the Agent Mesh UI, or feedback_enabled under spec.values for Slack and Teams.

  • For the Agent Mesh UI, confirm no platform-level override is set: call GET /api/v1/platform/webuiSettings and inspect collectFeedback. A null value means no override exists and frontend_collect_feedback applies; true or false means an override exists and takes precedence.

  • For the Agent Mesh UI, check the entrypoint logs for the message feedback configured but persistence not enabled; disabling for frontend, which is emitted when the UI requests its configuration.

  • Confirm session_service.database_url is set, then restart the entrypoint.

Feedback Events Do Not Reach Your Consumer

Publishing is enabled, but your subscribing service receives no feedback events.

  • Confirm publishing is on. Call GET /api/v1/platform/webuiSettings to see the platform override, which takes precedence over feedback_publishing.enabled in the entrypoint configuration whenever the override is set.

  • Confirm feedback collection is on: frontend_collect_feedback in the entrypoint configuration, or collectFeedback in the platform settings. With collection off, the Agent Mesh UI hides the rating buttons, so no rating is submitted and nothing is published.

  • Confirm your subscription matches {namespace}/sam/v1/feedback for the entrypoint's namespace. A {namespace}/sam/v1/feedback/> subscription does not match this topic, because > matches one or more further levels.

  • Confirm the entrypoint's event broker client is authorized to publish on that topic.

  • Verify the entrypoint's connection to the event mesh is healthy.

The API Returns 403 or an Empty List

A request to GET /api/v1/feedback returns 403 Forbidden or an empty data array.

  • A 403 on a taskId filter means the task belongs to another user. Each caller can retrieve only their own feedback.

  • An empty list means no stored feedback matches the filters for the calling user. Widen the startDate, endDate, or rating filters.

Feedback Is Not Pruned

Feedback older than the retention window remains in the database.

  • Confirm data_retention.enabled is true.

  • Confirm feedback_retention_days is set to the intended window.

  • The sweep runs every cleanup_interval_hours, which defaults to 24 hours, so allow one interval to pass before checking again.

What Next?

You have feedback collection, publishing, and retention configured. Most readers next want to see where feedback fits alongside the other operational signals Agent Mesh produces.