The problem

You want to analyse your helpdesk tickets — train a classifier, measure resolution times, share a sample with a vendor. The tickets are full of phone numbers, ID card numbers and bank details typed in by users, so you cannot.

Generic PII tools know about US Social Security numbers and miss a Moroccan CIN entirely. Tools that do match local formats tend to match too much: any 24-digit string becomes a "bank account", and your asset tags get shredded.

The approach: validate, don't just match

Every identifier that carries a check digit is verified before it is treated as personal data:

The effect: of 200 arbitrary 24-digit numbers, at least 190 are correctly left alone, because only about 1 in 97 can carry a valid key. Your inventory references survive; the bank details do not.

Installation

pip install glpi-anonymizer

Use it from the command line

# Straight redaction
glpi-anonymize tickets.json -o clean.json

# Keep records linkable without revealing anything
export GLPI_ANONYMIZER_KEY="a long random secret"
glpi-anonymize tickets.json -s pseudonymize -o clean.json --stats

# Leave a recognisable tail for support staff
glpi-anonymize tickets.json -s partial --keep 4 -o clean.json

It reads either a plain list of tickets or GLPI's {"data": [...]} envelope, and writes the same shape back. - reads stdin.

Use it from Python

from glpi_anonymizer import Anonymizer

anonymizer = Anonymizer()

result = anonymizer.anonymize("Call A123456 on 0612345678 or a.b@example.ma")
print(result.text)
# Call [CIN] on [PHONE] or [EMAIL]
print(result.counts_by_label())
# {<Label.CIN: 'CIN'>: 1, <Label.PHONE: 'PHONE'>: 1, <Label.EMAIL: 'EMAIL'>: 1}

clean_tickets, matches = anonymizer.anonymize_tickets(tickets)

Three replacement strategies

Pseudonymize is an HMAC of the normalized value, so the same number always yields the same token — across tickets, across files, across runs — while the original cannot be recovered. Normalisation means 06 12 34 56 78 and 0612345678 collapse to one token.

Keep the key secret. Identifier spaces are small enough to enumerate: anyone holding the key can confirm whether a given CIN is in your dataset.

Works with glpi-rest

Pull tickets and clean them in one pass:

from glpi_rest import GLPIClient
from glpi_anonymizer import Anonymizer

anonymizer = Anonymizer()
with GLPIClient(url, app_token=..., user_token=...) as glpi:
    tickets = glpi.get_items("Ticket", range="0-999")
    clean, matches = anonymizer.anonymize_tickets(tickets)

print(f"{len(matches)} values removed from {len(clean)} tickets")

How it works

graph LR
    IN["Ticket text<br/><i>Call A123456 on 0612345678</i>"]

    subgraph DETECT["1 · Detect"]
        PAT["Patterns<br/>CIN · GSM · RIB · ICE<br/>IBAN · card · email · IP · URL"]
        CHK["Check digits<br/>mod-97 · Luhn<br/><b>rejects look-alikes</b>"]
    end

    RES["2 · Resolve overlaps<br/>longest match wins<br/>RIB before ICE"]

    subgraph REPL["3 · Replace"]
        R1["Redact<br/>[PHONE]"]
        R2["Pseudonymize<br/>[PHONE:4f2a9c1b]"]
        R3["PartialMask<br/>******5678"]
    end

    OUT["Clean text + audit trail<br/><i>Call [CIN] on [PHONE]</i>"]

    IN --> PAT --> CHK --> RES
    RES --> R1 & R2 & R3 --> OUT

    style DETECT fill:#e8f0fe,stroke:#1a73e8,stroke-width:2px
    style REPL fill:#e6f4ea,stroke:#1e8e3e,stroke-width:2px
    style CHK fill:#fce8e6,stroke:#d93025,stroke-width:2px
    style OUT fill:#fef7e0,stroke:#f9ab00,stroke-width:2px

Overlaps are resolved before anything is replaced: the longest match wins, ties go to the higher-priority detector. That is why a 24-digit RIB is never carved up into a 15-digit "ICE".

Anonymizing an already-anonymized export is a no-op, so re-running the tool is safe.

Limitations — read before relying on this

This is a deterministic pattern matcher, not a language model. It does not detect:

• Names of people, which need named-entity recognition;
• addresses written in prose;
• personal details described rather than written as an identifier ("the manager on the third floor");
• identifiers split across lines or mangled by copy-paste.

Treat it as a strong first pass that removes structured identifiers reliably, not as a guarantee that a document is anonymous. Review a sample of the output before publishing anything. Under GDPR and Moroccan law 09-08, pseudonymized data is still personal data.

Security

No network access, no telemetry, no dependencies — the whole pipeline runs locally and the source is short enough to audit. See SECURITY.md.

Contributing

New detectors for other Maghreb identifiers are very welcome — see CONTRIBUTING.md.

git clone https://github.com/novagiosAI/glpi-anonymizer.git
cd glpi-anonymizer
python -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -e ".[dev]"
pytest

License

Apache License 2.0 — free for commercial use.