Quickstart

Save this as brief.md:

Mein Name ist Max Mustermann, erreichbar unter max@test.de.

privaparse --detector regex demo brief.md

demo runs the whole round trip and prints every stage — original, detections, pseudonymised text, a mock LLM answer, and the restored result. It is the fastest way to see whether the thing works on your documents. Person detection needs the [model] extra; without it, --detector regex limits detection to email and phone.

The real workflow is two commands, with your own trip to an LLM in between:

privaparse --detector regex pseudonymize brief.md -o brief.pseudo.md

Send brief.pseudo.md's contents to your LLM of choice and save its reply as antwort.md, then:

privaparse --detector regex reverse antwort.md -o antwort.klar.md

reverse with no --mapping looks up the mapping that issued every placeholder in the file. Partial coverage matches nothing, so this is convenience rather than a way around the mapping boundary — a file carrying a placeholder from a document you did not pseudonymise matches no mapping at all and is refused.

Pass --mapping <id> when you want to pin a specific mapping, and --mapping-out brief.id on pseudonymize to record the id at the time.

Other commands:

Command	Purpose
privaparse detect FILE --json	Show what would be detected; writes nothing
privaparse doctor	Resolved device, dtype, model, vault path
privaparse catalog show	Resolved catalogue — types, thresholds, sources
privaparse catalog validate [FILE]	Check a catalogue for errors; changes nothing
privaparse eval	Score detection against the gold set (needs GLiNER2)
privaparse bench	Throughput and detection quality together (needs GLiNER2)
privaparse vault stats	Counts only — never prints stored values
privaparse vault mappings	Recorded mappings and their ids, for a lost --mapping-out
privaparse serve	Run the gateway on 127.0.0.1
privaparse run -- <cmd>	Run a command with its OpenAI client pointed at the gateway
privaparse gateway stats	Counters from a running gateway — numbers only

As a library:

import privaparse

result = privaparse.pseudonymize(text)
answer = my_llm(result.text)
original = privaparse.reverse(result.mapping_id, answer).text

A long-running service should build one engine at startup instead, so the model is loaded once:

from privaparse.engine import PrivaParseEngine

engine = PrivaParseEngine()          # loads the model once
result = engine.pseudonymize(text)   # reuses it on every call