Skip to main content

​
Availability

Tyk AI Studio assigns a privacy score to LLM configurations, Tools, and Data Sources. This score controls which resources can be used together, so that sensitive data cannot reach a less trusted LLM.

​
The Privacy Score Scale

AI Studio stores the privacy score as an integer from 0 to 100. From v2.2.0, the UI shows the score as one of four named levels, together with the number: The create and edit forms for LLMs, Tools, and Data Sources have a level selector and a number field. When you select a level, the form sets its default score. You can also enter a different number in the range of the level. Lists, detail pages, the AI Portal, and submissions show the level and the number, for example Internal · 50. The API accepts only scores from 0 to 100. Other values return 400.

​
How the Comparison Works

A resource with a higher privacy score than the LLM it is used with cannot be attached to that LLM. For example:
  • A Restricted (100) Data Source cannot be used with an LLM Configuration whose privacy score is only Internal (50).
  • A Public (25) Tool can be used with any LLM Configuration, since 25 is the lowest score.
This check applies to:
  • Tools used with an LLM Configuration.
  • Data Sources used with an LLM Configuration in a RAG-enabled Chat.
  • Apps, when a user creates an App or an admin edits its resources. The check includes the Tools of the App.
  • Model Routers. A router has the lowest privacy score of the LLMs that it can reach. The router can send a request to any of these LLMs, so AI Studio treats the router as its least trusted LLM. For example, a router reaches an Internal (50) LLM and a Restricted (100) LLM. The score of this router is 50. An App with this router cannot have a Restricted data source. Otherwise, the router can send that data to the Internal LLM.
If a user tries to combine resources that fail this check, Tyk AI Studio blocks the action and shows an error.

​
Where You Set a Privacy Score

You set a privacy score when you create or edit: Community submissions to the AI Portal also carry a suggested privacy score from the submitter, which an admin can change before approving the submission.