Semantic Router Guard
The semantic router guard is two checks that the semantic router of a PaletteAI Inference Launchpad appliance runs on each request it reads to pick a category. This page explains the two checks, which requests each one applies to, and what happens when a check refuses a request. For which requests the semantic router reads, refer to What Reaches the Semantic Router.
The Two Checks
The guard makes two checks, and both are on by default. Each check has its own toggle on the Semantic router card under Settings > Configurations.
| Check | What it looks for | Which requests it applies to |
|---|---|---|
| Prompt attacks | Prompt attacks, such as jailbreak attempts | Every request the semantic router reads, whichever route the request would have taken. |
| Personal data | Personal data in the prompt | Only requests that would leave the appliance for a frontier provider or an external inference endpoint. |
Ordinary coding and general work does not fail either check.
When a Check Refuses a Request
A request read as a prompt attack is refused with HTTP 403, whichever route it would have taken.
A request that carries personal data and would leave the appliance is refused with HTTP 403, and nothing from it
leaves the appliance. A request that a local model answers is not held back for personal data.
Resources
-
Guard Toggles lists both states of each toggle.
-
Configure Semantic Routing walks through turning each check on or off.
-
Routing Behavior explains how the Tier map and the semantic router pick the model for each request.