Understanding the Methods
Your concept
Attribute Filter Value Matching
Two of the three methods, Foundation Model V1 and Analytics Agent V1, support attribute filter value matching. This means they work with the embeddings you configured in the Embedding Processing tab to match user language to actual data warehouse values. Custom Model V1 does not support this feature. If you select Custom Model V1, any embedding you configured will not be used during query generation. You need to specify exact attribute values in their requests for accurate filtering.
Choosing the Right Method
| Consideration | Recommended Method |
|---|---|
| You want the best overall accuracy for varied queries. | Foundation Model V1 |
| Speed is your top priority and queries are typically simple. | Analytics Agent V1 |
| Your workspace is English-only and you don't rely on embedding-based filtering. | Custom Model V1 |
| You use Embedding Processing for attribute filter matching. | Foundation Model V1 or Analytics Agent V1 |
Best Practices
- Use Foundation Model V1 as the default method. Foundation Model V1 provides a balanced combination of accuracy and capability for most use cases.
- Keep instructions short and specific. Each instruction increases processing cost and adds context that the AI must evaluate. Focus on rules that address observed issues.
- Use the Prompt Playground to test changes. After you add or modify instructions, test sample questions in the Prompt Playground and verify that the AI behaves as expected.
- Write instructions as rules, not as conversations. The AI interprets instructions as directives. State the required action clearly.
- Update instructions when business data changes. If the fiscal year changes, products are added, or terminology changes, update the instructions to match current information.
- Avoid conflicting instructions. Conflicting rules can produce inconsistent results. For example, one rule can filter by the current year, while another rule can include all years by default.
- Instructions apply to all BQL generation in the workspace. The instructions affect every user who uses AI capabilities in the workspace.
- Instructions do not override user permissions. The AI continues to enforce column-level access controls, even when instructions reference specific data.
- You can change Method Selection at any time. The change does not affect existing reports or configurations.
- Review model limitations before you change the method. When you switch from Foundation Model V1 or Analytics Agent V1 to Custom Model V1, attribute filter value matching is not available.