Voxta docs

Private data redactions

Strip secrets and private details from your messages, app context, documents and tool results before they reach the model.

Whatever a character reads goes to your model provider. A tool that reads a config file, a document you attach, a message you paste: any of them can carry an API key or a name you'd rather keep. Private Data Redactions replaces those before the model reads them.

It is part of Voxta Utilities: Text Replacements, a required module that is always installed, and runs in every chat while it has something to do.

What it covers

Text entering the chat from outside the model:

  • Your messages, and messages added by an app, a script or an augmentation.
  • Context set by an app or an augmentation.
  • Chat documents, title and content.
  • Tool results, errors and refusals included.

Never the model's own replies.

Settings

In the module's settings (Settings › AI Setup › Manage Modules, Voxta Utilities: Text Replacements), under Private Data Redactions:

  • Redact known secrets (on by default): private keys, API keys and tokens (OpenAI, Anthropic, AWS, GitHub, GitLab, Slack, Google, Stripe, Hugging Face…), JWTs, bearer tokens, passwords in URLs and card numbers. Each becomes [REDACTED api key], [REDACTED card number] and so on.
  • Custom matches: one regular expression per line. Add =>replacement to choose what it becomes, otherwise [REDACTED]. Lines starting with # are ignored.
(?i)\bJane Doe\b=>[name]
(?i)\bSpringfield\b
password=\S+

With both empty or off, nothing is redacted and the chat isn't told it has redactions.

Redaction matches shapes. A secret written in a form it doesn't know gets through, so keep path rules and tool review on as well. The server log notes that something was redacted, never the text itself.

On this page