> ## Documentation Index
> Fetch the complete documentation index at: https://phaseo.app/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Collect Feedback for Presets

> Record response ratings and compare preset quality in the Phaseo dashboard.

Preset Feedback lets you connect ratings from your application to the preset that produced a response. Use it to compare prompt and routing changes with real user signals instead of relying only on latency and cost.

<Note>
  Preset Feedback is in Alpha. We are actively refining the API and dashboard based on how teams evaluate production AI responses. Use the **Feedback** button on the [Preset Feedback page](https://phaseo.app/settings/presets/experiments) to tell us what would make it more useful.
</Note>

## Before you begin

You need:

* a Phaseo API key with the `feedback:write` capability
* the request ID returned by Phaseo
* the UUID of the preset you want to analyze

Include the preset ID when recording feedback. This is what makes the event appear in preset comparisons and cohort analysis.

## Record a rating

Send feedback after a user rates a response or your application determines an outcome.

```bash cURL theme={null}
curl https://api.phaseo.app/v1/feedback \
  -X POST \
  -H "Authorization: Bearer $PHASEO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "requestId": "gen_01K...",
    "presetId": "8f54b2d9-9ba3-4ced-8a7f-9bb70a44dbea",
    "rating": "thumbs_up",
    "score": 1,
    "comment": "The answer resolved the customer question.",
    "metadataDimensions": {
      "plan": "pro",
      "surface": "support_chat"
    }
  }'
```

A feedback event must target at least one request, session, preset, or test run. Scores are optional and must be between `0` and `1`.

## Choose a rating

Use a rating that describes the signal you actually collected:

* `thumbs_up` or `thumbs_down` for simple user reactions
* `correct`, `partly_correct`, or `incorrect` for evaluated answers
* `bad_format`, `too_slow`, or `too_expensive` for operational problems
* `unsafe` or `refused_incorrectly` for safety outcomes
* `not_helpful` or `other` when the other values do not fit

Avoid translating every signal into a numeric score. Explicit ratings make dashboard comparisons easier to explain.

## Compare presets

Open [Settings → Presets → Feedback](https://phaseo.app/settings/presets/experiments) to:

* compare positive, partial, and negative rates
* select a baseline preset
* filter by date, rating, and metadata dimensions
* inspect the request, session, comments, and tags behind an event

Add low-cardinality values such as plan, deployment, locale, or application surface to `metadataDimensions`. Keep arbitrary or high-cardinality context in `metadata` instead.

## Read feedback through the API

List recent events:

```bash theme={null}
curl "https://api.phaseo.app/v1/feedback?preset_id=PRESET_UUID&limit=100" \
  -H "Authorization: Bearer $PHASEO_API_KEY"
```

Request a server-side summary:

```bash theme={null}
curl "https://api.phaseo.app/v1/feedback/summary?group_by=preset&since=2026-08-01T00:00:00Z" \
  -H "Authorization: Bearer $PHASEO_API_KEY"
```

Read operations require `feedback:read`. Results are limited to the authenticated workspace and presets visible to the caller.

## Current Alpha limits

* The dashboard analyzes at most the 10,000 most recent matching feedback rows in a selected window.
* Preset comparisons use explicit ratings; numeric scores are shown as supporting detail.
* Test-run and custom outcome event APIs are available, but their dashboard workflows are still evolving.

## Related guides

* [Presets](./presets)
* [Roll out presets and debug routing](../cookbook/preset-rollout-and-routing-debug)
* [App attribution](./app-attribution)