Setting Up Embedding Processing

Enabling an Attribute

  1. Select Admin > Space Management > GenAI settings > Embedding Processing.
  2. Select the attribute you want to enable.
  3. Switch on the Enable option.
  4. Click the edit icon in the Embedding Model column.
  5. Select the model from the menu:
    • Adhoc: if the attribute has fewer than 100,000 unique values. No further steps required.
    • all-MiniLM-L6-v2: for English-only attributes with large value sets.
    • multilingual-e5-small: for multilingual attributes with large value sets.
  6. Optionally, expand Count of Unique Values to check the volume of data.
  7. Optionally, set a Top K value to control how many similar values the AI considers when generating filters.

Running Process Embedding

This step is only required for attributes using the all-MiniLM-L6-v2 or multilingual-e5-small models. Attributes using Adhoc skip this step entirely.
  1. After enabling attributes and selecting a processing model, click the Run Embedding Processing button in the top-right corner of the page.
  2. The system processes all unique values for each enabled attribute that requires it.
  3. Processing time depends on the number of unique values and the model selected.
  4. Once complete, the attribute is ready for use in AI-generated queries.

    When to run processing again:

    • After enabling a new attribute for embedding, with a non-Adhoc model.
    • After new data has been loaded or modified, for example after an ETL job.
    • After changing the embedding model for an attribute
    Note: The system only processes attributes that have changed or need updates. Previously processed values are not reprocessed unless the model configuration changes.

Configuring Top K

The Top K value determines how many similar attribute values the AI retrieves when matching a user's request. For example, if a user asks about "drinks" and Top K is set to 5, the AI will find up to 5 attribute values that are semantically closest to "drinks."
  • A higher Top K returns more potential matches, which can improve coverage but may include less relevant values.
  • A lower Top K returns fewer, more precise matches.
  • Leave this blank to use the system default.

Filtering and Searching

The toolbar at the top of the page provides options to manage large attribute lists:

  • Filter (All dropdown): Show only enabled attributes, disabled attributes, or all. Filter by model configured with the attributes.
  • Sort: Reorder the list alphabetically.
  • Search: Find a specific attribute by name.

Parent/Child Workspace Setup

If your workspace uses a parent/child configuration:

  • Embedding processing must be configured on the parent workspace, that is the workspace where the attribute data physically resides in the data warehouse.
  • Both parent and child workspaces must have GenAI enabled in their Space Properties.
  • Labels and descriptions for the same attributes are configured separately on the child (consumption) workspace.

Best Practices:

  • Start with the Adhoc model. It requires no processing step and works immediately. Only switch to another model if your attribute has more than 100,000 unique values.
  • Enable selectively. Only enable attributes that users are likely to filter by in natural language. Enabling too many attributes adds no benefit and may slow processing.
  • Re-run processing after data loads. If your data changes regularly through ETL, run embedding processing afterward to keep values in sync.
  • Test with the Prompt Playground. After configuring embeddings, use the Prompt Playground to verify that the AI is correctly matching user language to actual data values.
  • Review Top K settings if you notice the AI is including too many or too few filter values in generated queries.