Embedding Processing
By generating embeddings for attribute values, the AI can identify semantically related values and produce more accurate query results.
To access Embedding Processing, select .
Why use Embedding Processing?
When you ask a question that includes a filter, the AI must determine which values exist in the data warehouse. Without embeddings, the AI can only use the exact words provided in the user's question. If those words do not match stored values, the query may return incorrect or empty results.
Example:
You ask: "What are the sales of drinks products?"
| Without Embedding | With Embedding | |
|---|---|---|
| What happens | The AI filters for "drinks" directly from the user's question. | The AI searches for product category names that are similar to "drinks" in the data warehouse. |
| Generated query | SELECT [OrderDate: Sales] 'Sales' FROM [ALL] WHERE [Product.CategoryName] = 'drinks' | SELECT [OrderDate: Sales] 'Sales' FROM [ALL] WHERE [Product.CategoryName] IN ('coffee', 'soda') |
| Result | No data returned, "drinks" does not exist as a value in the warehouse. | Correct data returned, the AI matched "drinks" to semantically related product categories. |
Configuring Embedding Processing
The Embedding Processing page displays eligible attributes from your data model. For each attribute, you can enable embedding and configure how embeddings are generated and used.
For each attribute, you can configure these options:
- Enable
- Switch to activate embedding for this attribute.
- Name
- The attribute name and its parent dimension (read-only).
- Embedding Model
- The model used to generate embeddings for this attribute.
- Count of Unique Values
- Expand to view how many distinct values this attribute contains.
- Top K
- The number of closet matching results to return from the search.
Enable embedding only for attributes where:
- The values have meaningful, descriptive content, for example product names, categories, regions, departments.
- You are likely to ask questions that filter by these attributes.
- The values are something you might describe in different words than what's stored.
Examples of good attributes: Category Name, Product Name, Country, Department, Status.
Example of poor candidates: UUID, Employee ID, Phone Number, Email Address, Internal Codes.
Selecting an Embedding Model
When you enable an attribute for embedding, you must select an embedding model. Click the icon in the Embedding Model column to select from the available options:
| Model | Languages | Processing Required | Best For |
|---|---|---|---|
| Adhoc (Recommended) | 51 languages | No, it computes on demand at query time. | Attributes with fewer than 100,000 unique values. |
| all-MiniLM-L6-v2 | English only | Yes, it must run Process Embedding. | English-only attributes with more than 100,000 unique values. |
| multilingual-e5-small | 94 languages | Yes, it must run Process Embedding. | Non-English attributes with more than 100,000 unique values. |
The Adhoc model computes embeddings when a query runs. No separate processing step is required. After you enable the attribute and select Adhoc, the model is ready to use. Use Adhoc when the attribute contains fewer than 100,000 unique values. For attributes that contain more than 100,000 unique values, use a different model to prevent truncation.
Adhoc is the only available option for Live Access attributes.
All-MiniLM-L6-v2 is a compact, high-performance model for English-language attribute values. You need to run Process Embedding before you use this model. Select this model when attribute values contain English text and the attribute contains many unique values.
Live Access Space Attributes do not support this model.
Multilingual-e5-small is a compact embedding model that supports 94 languages. You need to run Process Embedding before you use multilingual-e5-small. Select multilingual-e5-small when attribute values contain non-English text and the attribute contains a large number of unique values.
Live Access Space Attributes do not support multilingual-e5-small.