# Scores and Feedback
Source: https://docs.genguardx.ai/agent-observability/scores-and-feedback/
Markdown: https://docs.genguardx.ai/agent-observability/scores-and-feedback/index.md
Description: Attach scores to LLM traces, spans and sessions in GGX: reviewer ratings, score configs, and end-user feedback relayed from your application with the SDK or the API.

A **score** is a judgment attached to a trace, to one span, or to a whole session: a reviewer's rating, a cardholder's thumbs-down relayed by your application, an evaluator's verdict, or a score imported from Langfuse. Every score records who gave it, when and from where, and a changed score keeps its history.

| Source | Who writes it | Can be edited by |
| --- | --- | --- |
| Reviewer | Someone scoring from the trace page | Its author, or an organization admin |
| API | Your application or a script | Its author, or an organization admin |
| End user | Your application, relaying its user's feedback | Its author, or an organization admin |
| Judge | An [evaluator](https://docs.genguardx.ai/agent-observability/evaluators/) | Nobody. An organization admin may delete it. |
| Imported | The Langfuse import | Nobody. It is someone else's record. |

## Scoring from the trace page

1. Choose **Scores & evaluations** on the trace, then **Add score**, to score the whole trace. To score one step instead, open that span's **Scores** tab.
2. Pick a score name from your project's score configs, or type a new one.
3. Enter a number, yes/no or label, and add a comment.

The trace list shows each trace's scores and can filter on them.

## Score configs

A score config fixes what a score name means in a project, so two reviewers' _helpfulness_ is the same measurement. It sets the type (number, yes/no or label), the valid range or labels, and guidance for reviewers. A score whose name has a config must fit it.

Manage them in the **Score configs** tab.

*Figure: The Score configs tab listing a yes/no score and a 1-to-5 numeric score with reviewer guidance* Archiving a config hides it from pickers but keeps enforcing it.

## Relaying end-user feedback

Your application can score the exact trace a user is reacting to, straight after the call and before the trace has even finished arriving.

```python
from genguardx import tracing

trace_id = tracing.current_trace_id()   # read it inside the call being rated
...
tracing.score(
    trace_id, "thumbs", False,
    comment="blocked the wrong card",
    source="end_user",
    external_id=request_id,              # a retry with the same id updates, never duplicates
)
```

A deployment called through `POST /api/v1/deployments/invoke` returns `traceId` and `projectId` in its response, and an `X-Trace-Id` header, for the same purpose. Without the SDK, post the score to [`POST /api/v1/trace-scores`](https://docs.genguardx.ai/agent-observability/api-reference/#score-request).

**Note**

`tracing.score()` raises `tracing.ScoreError` when GGX refuses a score, for example a value outside the score config, or when tracing is not initialized. Unlike sending spans, a score you asked for and did not get is something you need to know about. Comments are stored as written, so keep personal data out of them.

## What to read next

- [Evaluators](https://docs.genguardx.ai/agent-observability/evaluators/): judges that write scores for you, on live traffic.
- [Monitors](https://docs.genguardx.ai/agent-observability/monitors/#what-you-can-monitor): alert on a score's mean or the share of a label.
- [Search, analytics and export](https://docs.genguardx.ai/agent-observability/search-and-analytics/#the-query-box): query traces by `score.<name>`.