> For the complete documentation index, see [llms.txt](https://docs.radar.citadel-ai.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.radar.citadel-ai.com/en/application-owner-guide/evaluation-library.md).

# Manage the Evaluation Library

In the Evaluation Library, you can manage metric-related resources for evaluating LLM applications.

By associating Evaluation Metrics and Evaluation Packages created in the Evaluation Library with an application, you can evaluate input and output logs and detect risks in AI systems.

This page explains the relationship between Evaluation Categories and Evaluation Metrics as a basic concept for understanding the Evaluation Library.

### What is an Evaluation Category?

An Evaluation Category is a classification for grouping related Evaluation Metrics.

For example, an Evaluation Category such as `Disclosure of Organizational Confidential Information` includes evaluation metrics that target inputs and evaluation metrics that target outputs. Each evaluation metric is version-managed.

By reviewing the Evaluation Categories, you can understand what kind of risk domains are covered by each of the Evaluation Metrics.

#### Note: relationship between Evaluation Categories and Evaluation Metrics

{% hint style="success" %}
Evaluation Categories are units that define risk and threat domains. The targets of registered evaluation metrics differ by Evaluation Category.

**Cases where risks or threats are included in the inputs**\
These are cases where the user input itself contains risky information or instructions.\
For example, in the `Input of Personal Information` category, only metrics that target inputs are registered in order to evaluate whether user input contains sensitive personal information.

**Cases where risks or threats are included in the outputs**\
These are cases where the LLM output contains the content that should be detected.\
In addition to whether problematic output has been generated, the input that attempts to elicit that output is also evaluated.\
For example, in the `Disclosure of Personal Information` category, input metrics that evaluate questions or instructions intended to obtain or infer personal information and output metrics that evaluate whether output containing actual personal information has been generated are registered.
{% endhint %}

### What is an Evaluation Metric?

An Evaluation Metric is an evaluation criterion that Citadel Radar uses to detect risk.

There are mainly two types of Evaluation Metrics.

<table><thead><tr><th width="197.90625">Metric type</th><th>Overview</th></tr></thead><tbody><tr><td>Built-In Metrics</td><td>Evaluation Metrics pre-installed in Radar. They can be used as-is for common risk domains.</td></tr><tr><td>Custom Metrics</td><td>Evaluation Metrics that can be created to match company-specific evaluation criteria, guidelines, rules, and similar requirements.</td></tr></tbody></table>

For details about Built-In metrics, see [Review Built-In Metrics](/en/application-owner-guide/evaluation-library/built-in-metrics.md).

For how to create Custom Metrics, see [Create Custom Metrics](/en/application-owner-guide/evaluation-library/custom-metrics.md).