Performance and latency

Anonymization sits on the critical path between your users and the model, so its cost matters. This page separates the two sides, because they behave very differently.

De-anonymization is negligible

Restoring the real values in a reply is a single pass over the text: every token of the fixed shape is resolved against the session mapping in one lookup, independent of how many tokens there are. On a streaming (SSE) reply, where restoration runs on every delta, this was measured at about 0.12s across a 50-delta stream. It does not degrade the interactive feel.

Detection scales with input size

The cost is on the detection side, and it is NLP bound, so it grows with the length of the text being scanned rather than with a fixed per-request penalty.

• The NLP models are loaded once at startup, so there is no per-request cold start. Detection on a short prompt is in the low tens of milliseconds once warm.
• Deterministic detectors (emails, IPs, IBANs, card numbers, national identifiers) are cheap: they are pattern and checksum checks.
• The heavier part is named-entity recognition for names, organizations and locations in free text. A large context, such as whole files pasted into a prompt, adds detection time proportional to its length.

Realistic interactive workloads

For an interactive developer or chat workload, the dominant cost is the model call itself, not anonymization. Detection runs on the outbound content before that call, and restoration runs on the streamed reply at negligible cost.

The honest position on exact figures: they depend on your content mix (how much free-text prose versus structured identifiers) and your hardware. Rather than quote a single number, we benchmark on your realistic contexts during a pilot and give you hard measurements for your own workloads.