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

# Evaluating a Request

> Run one-off evaluations on a single request (and compare versions)

The **Request Overview** page includes an **Evaluation** section that lets you inspect existing evaluations and run additional one-off evaluations using your Sentinels.

See: [Request Overview](/observe/request_overview)

## When this is useful

Evaluating a single request is great for:

* Debugging a **specific failure mode**
* Verifying whether a proposed Sentinel would have flagged a request
* Comparing evaluation outcomes between the **original request** and a **rerun**

## What you’ll see on the Request page

The Evaluation section shows one or more **evaluation blocks**. Each block includes:

* The set of **selected sentinels**
* A request-level summary (pass count / total)
* Per-sentinel results (status, description, eval time)

Blocks can represent different sources of evaluation:

* **Real-time**: evaluations run automatically when the request was originally served
* **Batch Eval**: results from a batch evaluation run over a date range of requests
* **Manual**: one-off evaluations you run yourself from the Request page

## Run a manual (one-off) evaluation

1. Open a request and scroll to the **Evaluation** section.
2. Click **Add Evaluation** to create a new manual evaluation block.
3. Select one or more sentinels to run.
4. Click **Run Evaluation**.

<Note>
  You can only run an evaluation once the request has a final **assistant** message to evaluate. If you want to try a different sentinel set after running, add a new evaluation block.
</Note>

## Evaluating different versions (original vs reruns)

If you rerun a request (for example, with a different model or prompt), you can add an evaluation block while that version is selected to evaluate the rerun output with the same sentinels.

## Suggested corrections

If a sentinel returns **FAULT**, the UI may show a **Suggested Correction** when the evaluation result includes one. Treat these as guidance for debugging and refinement, not as an automatic change to the request.