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

# Create evaluators

> Use the evaluator library and build your own custom evaluator.

Evaluators turn a model's free-text responses into numbers you can compare and aggregate. The library includes rule-based and statistical evaluators, marked **Default**, and you can create your own LLM-as-a-Judge evaluators to score against criteria you write.

You can author an evaluator with nothing else set up, but testing it or running it in an examination needs at least one registered [model](/guard/app/models).

## Browse the evaluator library

The library is a table, with **Name**, **Description**, **Type**, **Created By**, **Last Updated**, **Visibility**, and **Actions**. Evaluators SeekrGuard provides show a **Default** badge under **Created By**.

Above the table, filter by **Name**, **Type**, or **Model**, reorder the list with **Sort By**, and limit it to your own evaluators with **Only mine**.

Each row's **Actions** column offers:

* **View Evaluator** – open the evaluator's detail page.
* **Delete Evaluator** – available only for evaluators you created.

## Create an evaluator

The creation flow builds LLM-as-a-Judge evaluators. The rule-based and statistical evaluators come with SeekrGuard and cannot be created.

<Warning>
  An evaluator cannot be changed after you create it.
</Warning>

<Steps>
  <Step title="Add Evaluator">
    On the **Evaluators** page, click **Add Evaluator**.
  </Step>

  <Step title="Configuration">
    * **Name** – a short label for the evaluator.
    * **Description** – the judge model relies on this to decide how to evaluate and score outputs. Write it as instructions rather than a label.
    * **Evaluator Model** – the model that acts as the judge. Optional here, since you pick judge models when you add the evaluator to an examination.
    * **Scoring Scale** – **Minimum**, **Maximum**, **Step**, and **Polarity**. The defaults are 0 to 5 in steps of 1, with positive polarity.
    * **Constraints (Optional)** – natural-language rules for the judge to apply.

    <Warning>
      Set polarity to match your metric. A metric where high values are bad, such as toxicity, must be negative. Risk scoring inverts non-positive polarity so that higher always means more risk, and a wrong setting makes a model's risk scores read backwards. See [Polarity](/guard/glossary#polarity).
    </Warning>
  </Step>

  <Step title="Examples (Optional)">
    Add few-shot examples to calibrate the judge. Each has **Text**, a **Score**, and optional **Reasoning**.
  </Step>

  <Step title="Preview & Submit">
    Review the **Name**, **Model**, and the **Judge Prompt** SeekrGuard generates from your description and constraints, then click **Create Evaluator**.
  </Step>
</Steps>

## Read an evaluator's detail page

**View Evaluator** opens a page showing the evaluator's **Description**, its **Inputs**, and **Facts** such as type, dates, score range, and polarity. **Inputs** lists what the evaluator needs in order to run, such as the model's output, the expected answer, or retrieved context. Those requirements determine the datasets it can be used with. See [Run examinations](/guard/app/examinations).

The **Test Evaluator** panel scores text you supply, without creating an examination. Choose a judge model with **Select Model**, which lists the judge models this evaluator has been used with. Expand **Judge Prompt** to read the full prompt the judge model receives. Fill in each input the evaluator requires, such as **LLM Output**, then click **Run Test**. **Load Example** fills the fields with sample text and **Clear** empties them.