Why

You need ticket data — to demo a plugin, load-test an instance, or train a classifier. Real tickets are full of personal data and usually cannot leave the building, which stops the project before it starts.

The usual workaround is a loop that writes "Test ticket 1, Test ticket 2". That tells you nothing: no priority distribution, no escalations, no language mix, nothing a model or a dashboard could learn from.

This package generates tickets that behave like real ones, and that contain no personal data by construction — there is no source dataset to leak from.

Installation

pip install glpi-testkit

Use it

# 1000 tickets, reproducible, with a distribution summary
glpi-testkit -n 1000 --seed 42 --stats -o tickets.json

# Darija only, VPN incidents only
glpi-testkit -n 200 -l ary -c network.vpn -o darija-vpn.json

# Ready to POST to the GLPI API
glpi-testkit -n 50 --glpi-input -o payload.json

# See the catalogue
glpi-testkit --list-categories

from glpi_testkit import TicketGenerator, Language

generator = TicketGenerator(seed=42, languages=[Language.FRENCH, Language.DARIJA])
tickets = generator.generate(1000)

print(tickets[0].name, tickets[0].priority.name, tickets[0].language)
print(tickets[0].to_glpi_input())  # ready for the GLPI REST API

What makes the data usable

It follows GLPI's own priority rules

GLPI does not average urgency and impact — it looks the pair up in a configurable 5×5 matrix. The default below was read out of a stock GLPI 11 install, and the arithmetic fallback in CommonITILObject::computePriority is reproduced too, so a ticket generated here carries the priority a real GLPI would have computed.

         impact:  1  2  3  4  5
urgency 1:        1  1  2  2  2
urgency 2:        1  2  2  3  3
urgency 3:        2  2  3  4  4
urgency 4:        2  3  4  4  5
urgency 5:        2  3  4  5  5

from glpi_testkit import compute_priority

compute_priority(urgency=4, impact=5)  # Level.VERY_HIGH

Pass your own matrix if you have customised yours: TicketGenerator(priority_matrix=my_matrix).

The lifecycle holds together

Every ticket carries a coherent history, which is what makes the data worth training on: timestamps are ordered, only assigned tickets have an assignee, only solved ones have a solve date, delay statistics match their dates, and high-priority tickets get escalated and bounce between groups.

stateDiagram-v2
    [*] --> NEW
    NEW --> ASSIGNED: picked up
    ASSIGNED --> ASSIGNED: escalated / reassigned
    ASSIGNED --> WAITING: awaiting requester
    ASSIGNED --> PLANNED: scheduled
    WAITING --> ASSIGNED
    ASSIGNED --> SOLVED: solution provided
    SOLVED --> CLOSED: validated
    CLOSED --> [*]

    note right of NEW
        4% stay here
    end note
    note right of CLOSED
        55% end here,
        as in a real export
    end note

SLA pressure is a dial, not an accident

sla_pressure scales resolution time against each category's target. Measured over 3000 tickets across five seeds:

The three languages are written properly

French, Modern Standard Arabic, and Darija in both Arabic script and Latin letters (arabizi) — because that is what a Moroccan service desk actually receives.

fr   Le scanner de l'agence de Rabat est hors service depuis deux semaines.
ar   تظهر الطابعة رسالة انحشار الورق رغم أن الدرج فارغ.
ary  l vpn kat9ta3 kol chi 10 d daqayq, tan7iyed men SAP kol mrra.

Slot substitution is the easy way to get variety and the easy way to produce broken grammar — "un imprimante", "du agence", "de le CRM", or an Arabic feminine verb against a masculine noun. Every slot therefore carries its own article, preposition and gender, and a dedicated test suite scans thousands of generated sentences for each fault class.

Darija written in Latin letters is deliberately left lowercase: users really do type it that way, and that is what a classifier will meet in production.

Native review welcome. The Arabic and Darija phrasings are written to be idiomatic, but a native reviewer will always improve them — corrections are the single most useful pull request this project can receive.

Load it into GLPI

With glpi-rest:

from glpi_rest import GLPIClient
from glpi_testkit import TicketGenerator

tickets = TicketGenerator(seed=42).generate(200)

with GLPIClient(url, app_token=..., user_token=...) as glpi:
    for ticket in tickets:
        glpi.add_item("Ticket", ticket.to_glpi_input())

to_glpi_input() emits only the fields the API accepts on creation, with GLPI's integer enums and Y-m-d H:i:s timestamps. to_dict() gives you everything instead — events, delay statistics, requester and assignee — for offline work.

The catalogue

Urgency and impact ranges are set per category, so a jammed printer never outranks a security incident by chance.

Privacy

There is no source dataset. Every name, phrase and reference in this package was invented, and a test asserts the output contains nothing shaped like a Moroccan phone number, RIB or ICE.

That is the point: it lets you build and publish on ticket data without ever processing a customer's. If you also need to clean a real export, see glpi-anonymizer.

Limitations

• Seven categories today. Real service desks have dozens — new catalogue entries are the most useful contribution.
• Text is template-based, so lexical variety is bounded by the catalogue. It is built for structure, volume and language mix, not for fooling a human reader into thinking a person wrote it.
• Timestamps ignore business hours and public holidays.

Contributing

See CONTRIBUTING.md.

git clone https://github.com/novagiosAI/glpi-testkit.git
cd glpi-testkit
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.