DocumentationAPI ReferenceRelease Notes
DocumentationLog In
Documentation

User-Value Properties

Learn what types of user-value properties exist in Deepchecks LLM Evaluation and how they are defined

User-Value Properties are any numerical or categorical properties, sent to Deepchecks by the user, that provide additional context or metrics about the interactions.

Once added to the system, User-Value properties can be utilized like any other property. They are available for a wide range of purposes, including annotation, root cause analysis, version comparison, and production monitoring, thus enhancing the depth and versatility of your evaluations.

Types of User-Value Properties

User-Value properties can be broadly categorized into two types: metadata properties and metric properties.

User Value Icon

User-Value Icon

Metadata Properties

Metadata properties are descriptive attributes that provide contextual information about a sample. These properties help in organizing and categorizing samples based on various descriptive factors.

  • An example for a metadata property could be the "Region" associated with a sample, such as "North America", "Europe", etc.
  • In the system, metadata properties are used as part of the Root Cause Analysis (RCA). Users can receive insights and reports based on custom properties categorization.

Metric Properties

Metric properties are quantitative values that assess certain aspects of the sample. These properties provide measurable data that can be used for evaluation and comparison.

  • An example for a metric property could be "Internal Document Quality" representing the vector store retrieval score.
  • In the system, metric properties can be integrated into the Automatic Annotation flow and affect the estimated scores provided by Deepchecks.
  • You may provide reasons for any given value. For example, you may want to add a reason of "bad grammar" to a low value of the "Internal Document Quality" property for a specific interaction. These reasons can then be viewed in the interaction view in the system.

By leveraging both metadata and metric properties, users can tailor the system to better meet their specific needs, improving the overall management and analysis of samples.

Creating a New Custom User-Value Property

New properties have to be defined as "Categorical" or "Numeric". This can not be edited later.
Once the User Value property was created, you can start accommodating it with values via our SDK as well as in the UI with CSV Upload, by adding a value for the corresponding property for each interaction.

📘 Note: Defining User-Value Properties before sending or uploading their values

The User-Value property should be created for the interaction type before supplying the custom property values for those interactions, otherwise they won't be identified and saved by the system.