LLM providers
By default, Omni’s AI features use LLMs hosted on Amazon Bedrock for AWS deployments, or Microsoft Foundry for Azure deployments. Refer to AI inference regions for more information. Organizations can also configure custom AI model providers, including Anthropic Direct, Google Vertex AI, OpenAI, and Grok (xAI). Refer to the alternative model providers documentation for more information about configuring custom providers. Organization Admins control which LLMs their organization can use. Refer to the LLM settings documentation for more information. Any data, including metadata, provided to the LLMs used by Omni will not be used for training. When using custom providers, API keys are securely stored per organization and are not exposed in the UI after saving.User prompt categorization
When a user enters a prompt, Omni will first use the LLM to categorize the request to determine how to respond.Shared data
To determine the correct category for the prompt, Omni shares the prompt itself with the LLM. Specifically, this is the natural language question or prompt provided by the user. For example, “How many users signed up last month?” or “Filter by the last two years.”Query generation
When generating a query, the output of the AI’s processing is a new Omni query, which is a collection of metadata - including field names, filters, sorts, pivots, topic name, and limit - that is translated into SQL. This new query is then run within the current workbook to provide an answer to the user’s prompt. Because the AI uses an abstract query format and not SQL to create queries, the generated query respects the permissions set in Omni. This means it’s not possible to access data outside of the topics, models, and connections a user has been restricted to in Omni. Additionally, this approach ensures that no private, relational, or result set data is shared to generate queries. Only metadata about the current query, user prompt, and selected topic fields is sent to the LLM, thus maintaining the privacy and security of your data.Shared data
While relational data is not shared to generate queries, Omni does share certain metadata with the LLM to generate accurate responses. This metadata includes:| Description | What’s shared? |
|---|---|
| Current query metadata | Information about the currently selected query in the workbook, such as field names, sorts, limits, pivots, and filters |
| User prompt | The natural language question or prompt provided by the user. For example, “How many users signed up last month?” or “Filter by the last two years.” |
Context (ai_context) | Free text that provides context. This is set at the topic or view level using the ai_context parameter. |
| Fields within the selected topic | Metadata about the fields in the currently selected topic, such as field names, labels, descriptions, data types, and whether they’re aggregates (count, sum, average) or dimensions. Note: Even if a user has access to multiple topics in a workbook, only metadata for the topic that’s currently selected is accessed. |
Data summarization
Omni’s AI uses the LLM to power its data summarization features, which includes using the AI summary visualization or the Omni Agent to summarize the results of a query.Shared data
The region where AI inference is performed depends on where your Omni instance is hosted, and the inference region isn’t always the same as your hosting region. Refer to AI inference regions for more information.
- The metadata for the current query
- A CSV of the current query’s results
Web search
Web search is only available for Omni instances hosted on AWS in a supported cloud region. See Cloud regions for more information.
Shared data
When web search is enabled, the following data is shared with AWS to perform the search:- The search query text
- Any domain or date filters specified for the search