European PII Redaction.
Open-source SDK that detects and redacts structured PII across 31 European countries. Available for Python and Node.js. Deterministic, and local in its default rules mode: no network calls. Only cloud mode contacts euRedact Cloud.
On the published benchmark, euRedact Cloud leaves 7% of personal data readable; the average over the five systems measured is 35.5%. The benchmark →
Redaction in 30 Seconds.
$ pip install euredactimport euredact text = "Mijn BSN is 111222333 en IBAN NL91ABNA0417164300." result = euredact.redact(text, countries=["NL"]) print(result.redacted_text)
* euRedact Core, measured on a generated evaluation set with countries supplied — the conditions are in the measurement note. † euRedact Cloud on the published 900-document benchmark — every type, set and language on the cloud benchmark page, the conditions below.
Live Demo
Demo for professionals only. euRedact Cloud is offered to businesses and professionals; this demo is for evaluating it in that capacity, not for consumers. See the demo notice.
euredact · PyPIThe European PII Problem
Cloud API Risks
Sending documents to a third-party redaction API means personal data leaves your infrastructure — creating a new processor relationship and transfer surface under the GDPR. euRedact Core avoids this; euRedact Cloud reduces it: with the SDK, the identifiers the rules engine finds are masked on your machine first, and the rest is processed in the EU under a DPA.
Localization Gaps
Tools built for US data struggle on European documents. Identifiers such as the BSN, codice fiscale and NHS number, and the national ID and IBAN formats used across 31 countries, need per-country rules and checksums.
Slow Development
Building and maintaining per-country detection rules in-house is slow and hard to verify.
Processing Architecture
Steps 1, 2 and 4 run on your infrastructure. Step 3 runs on euRedact Cloud in the EU and receives the text after the rules engine has masked what it recognises. The rules engine is deterministic; the same input always produces the same redaction.
Input
Betaling aan Jan de Vries, BSN 111222333, vanaf NL91ABNA0417164300
Local Engine
Deterministic rules resolve structured identifiers per country, confirmed by checksum.
31 country rulesetsAI Model
Public betaContextual detection for names, addresses and other unstructured PII the rules engine does not emit.
Redacted Output
Betaling aan Jan de Vries, BSN [NATIONAL_ID], vanaf [BANK_ACCOUNT]
The name is left in place — it needs the cloud tier at step 3.
The Hardest Categories
The GDPR’s Article 9 and 10 data — health, religion, political opinions, offences — rarely has a shape, does the most harm when it leaks, and is where we put our effort release after release. On a published benchmark of 900 documents in five languages, that shows: F2 0.921 overall against a benchmark average of 0.695 over the five systems measured, euRedact included — and the gap is widest exactly here.
Measured on 14 November 2026 through the production service with SDK 1.0.0, on three synthetic sets of IT, HR and medical documents in English, French, German, Dutch and Polish (27,831 annotated items, none in euRedact’s training data); the four other systems on 2 November. “Readable” means less than 75% of an item’s letters and digits hidden; “masked” is the opposite; “found” counts an item when any part of it was redacted. F2 counts recall twice, because a missed item leaks. Every type, set and language is on the cloud benchmark page; the method and the other systems’ set-ups are in the launch post; Presidio and euRedact side by side, on this benchmark and on European identifiers, on the comparison page and the benchmark on GitHub.
Deployment Options
euRedact Core
Apache 2.0The free, open-source SDK. Local, deterministic, benchmarked.
Nothing leaves your machine. In rules mode the SDK makes no network calls; the identifiers it finds are masked where the document is.
- Deterministic rules engine — same input, same output
- 31 European country rulesets, 29 entity types
- 373 pattern definitions, 48 validators
- Python and Node.js SDKs, zero dependencies
- Runs locally — rules mode makes no network calls
euRedact Cloud
Public betaManaged, EU-hosted neural layer for contextual and unstructured PII.
Nothing is kept. By default, text and results are deleted as soon as the result is delivered; only content-free usage statistics remain.
- Everything ineuRedact Coreplus:
- Neural detection for unstructured and contextual PII
- 11 further types: names, addresses, places, medical information
- Managed, EU-hosted API
- With the SDK, identifiers the rules engine recognises are masked on your machine before text is sent
- Public beta: a free plan now, paid plans from 14 December
Built for Europe
31 countries with maintained detection rules for national identifiers, tax numbers, and local formats.
The rules engine ships today across all 31 countries. Cloud coverage describes the contextual model in public beta, which handles Dutch, English, French, German and Polish; a country counts as fully covered when every language its government and business documents are written in is among those. Minority co-official languages are out of scope.