The round-trip workflow

Deanonymization exists for one pattern: safely using a tool you don’t control on documents you can’t share. Anonymize on the way out, restore on the way back. Real PII never crosses the boundary.

The loop

Your document  ──anonymize──▶  safe copy  ──▶  external tool (LLM, colleague, service)
      ▲                                                   │
      └──────────── deanonymize ◀──── edited/answered copy ┘

1. Anonymize with synthetic mode (so the document still reads naturally) and a dictionary attached with add to dictionary on (so the swaps are recorded).
2. Use the external tool on the safe copy. It sees only fake values — David Romero Gil, not Marcos Patel.
3. Bring the result back — the tool’s edited file, summary, or answer, still written in fake values.
4. Deanonymize it. Piixie swaps the fakes back to the originals, exactly if it recognizes the file, or via the dictionary if the tool changed it.

A worked example: summarizing a medical record with a cloud LLM

You need a cloud model to summarize a patient’s history, but the record can’t leave compliant systems as-is.

1. Anonymize expediente-clinico-marcos-patel.pdf → expediente-clinico-david-romero.pdf, synthetic, Español, with your Clínica dictionary growing.
2. Paste the safe copy into the cloud chat: “Summarize this patient’s cardiac history.” The model answers about David Romero Gil, NHC 84913366.
3. Save the model’s answer to a text file.
4. Drop it into Piixie with the Deanonymize toggle on. The file wasn’t produced by Piixie, so it uses the dictionary: David Romero Gil → Marcos Patel, 84913366 → 1029384. You get a summary about the real patient, having never sent the real patient to the cloud.

The “paste safely, then restore” flow is also covered, end to end, in Share safely with an LLM and restore.

Why synthetic mode for the round trip

Only synthetic and labeled output is reversible — they preserve a swap you can undo. Redaction destroys the original. And synthetic specifically keeps the document readable, so the external tool does a good job on it: a model summarizes a natural document far better than a sea of [REDACTED].

Keep the means to reverse

The round trip only closes if Piixie can get back. Two belts, one suspenders:

• Keep the output file intact if you can — that enables exact reverse.
• Grow a dictionary from your runs — that enables dictionary reverse even when the file was changed downstream.

With both, you can restore whether you get back the same file or a rewritten one.

The boundary, restated

The restored document has real PII again — it’s back on the trusted side. Treat the trip as: plaintext in (yours) → fake out (theirs) → fake in (theirs) → plaintext out (yours). The external side only ever holds fakes. That’s the guarantee. More on what’s stored and what crosses the network: privacy.

Next

• Limits & safety
• Share safely with an LLM and restore