I’m trying to understand how Perplexity performs real-time web searches while avoiding the exposure of sensitive user data to the public internet.
From what I assume , it looks like the data flow might include several layers of protection:
- A prompt filter for PII or sensitive terms.
- Search Agents then operate in isolated containers to prevent data leakage.
- Execution of live web search is done with abstracted keyword-based searches, instead of raw prompts.
- Results of websites are then temporarily cached and then answers aggregated before being returned to the user back via the API gateway
So my question is:
If a prompt contains sensitive or proprietary information, how exactly is that protected from being sent out over the public internet during web search execution?
Is the abstraction purely semantic or keyword-based? Does anything ever get logged or exposed beyond internal systems when the agents do live web searches?
Any clarity on how the abstraction layer for Sonar Pro and Sonar Deep Research API’s and their isolation mechanisms work in practice would be super helpful to understand how the proprietary data in a prompt is never publicly exposed.