Skip to content

Adding a missed entry

No detector is perfect. When you spot a value the model missed, you don’t re-run anything — you select it and add it as a new entry. From then on it’s treated like any other.

  • Select, then click +. Highlight the text in the viewer, then click the + button in the entries header. The selection is pre-filled.
  • Click + first. Open the dialog empty and type the original yourself.

The Add entry dialog opens with the original text, a type, and a method.

  • Original — the text to match. Pre-filled from your selection; editable.
  • Type — pick from the catalog. This sets the tag and steers synthetic generation.
  • Method — four choices:

A realistic fake value. Pick a language / region and optional gender, then click Replacement to preview a generated value. Example: type EMAIL, region Español[email protected].

A masking fill. Pick a redaction character (or a custom one) and the value becomes that character repeated to length.

A fixed string you type. Use it for a known mapping — Proyecto ORION should always become Proyecto AZUL in this document.

A tiny JavaScript function that returns the replacement. The default skeleton:

function value(original, type) {
// return the replacement for this entry
return "REDACTED";
}

This is the escape hatch for a rule a generator can’t express — derive a code from the original, look up a convention, format something just so. It runs in Piixie’s embedded JavaScript engine, the same one profiles use.

The new entry lands at the top of the list, marked as yours, so method switches won’t overwrite it. If the same text appears several times, the new entry covers every occurrence — check the occurrence badge to confirm the count.

Codenames, internal IDs, and the long tail

Section titled “Codenames, internal IDs, and the long tail”

The values most worth adding by hand are the ones a general model has no way to know are sensitive: an internal project codename, a customer reference format unique to your company, a building nickname. Add them once here, and if they recur across documents, promote them into a dictionary or bake the rule into a profile prompt so future runs catch them automatically.