Skip to main content

Interceptors

An interceptor is a guardrail applied to the document's text before it is sent to an AI agent. It acts on the whole document or on each segment passed to the analysis (including for each agent driven by a supervisor).

Matching types​

  • Pattern (regex) — a standard regular expression. Any custom expression is automatically tested against load scenarios before being accepted, to prevent a poorly written regex from blocking processing (denial-of-service protection).
  • AI judgment — a natural-language instruction evaluated by a dedicated AI model, which returns the list of passages to flag.
  • Decision model — a yes/no question (e.g. "Does this text contain health data?") asked to a decision model. Processing is blocked as soon as the probability of "yes" reaches the blocking threshold (50% by default). This type can only block: the model answers with a probability, without pointing at passages to mask. A text longer than the model's window is judged piece by piece.

Available actions​

  • Mask — replaces each detected passage with a substitution text (<REDACTED> by default); the analysis then continues normally on the modified text.
  • Block — immediately stops the agent's analysis as soon as the first match is found.

Built-in interceptors​

Five personal-data detection interceptors are provided by default (disabled by default): emails, credit card numbers, IP addresses, MAC addresses and URLs. For these built-in interceptors, only the action (mask/block/disabled) can be changed — their detection pattern cannot be modified, and they cannot be deleted.

Configuration​

Reading requires Interceptors — read; creating, editing and deleting custom interceptors requires Interceptors — write.