What it does

$ ticket-triage predict "l vpn kat9ta3, ghir l pc dyali, machi musta3jel"
category : network.vpn  (60%)
priority : 1  (58%)

$ ticket-triage predict "تظهر الطابعة رسالة انحشار الورق. جميع الموظفين متوقفون. الأمر عاجل"
category : hardware.printer  (79%)
priority : 5  (55%)

Same topic, opposite priorities — and the model got there by reading "only my PC, not urgent" versus "all staff are stopped, it's urgent", in two different languages. That is the whole trick, and the section below shows how it was measured rather than asserted.

Results

Trained on 9000 synthetic tickets, tested on 1430 held-out ones:

Read those numbers with the three caveats below, not without them.

1. The category score is not an achievement

100% looks like a result and is not. The seven categories use disjoint vocabulary by construction — VPN, imprimante, mot de passe — so a linear model separates them perfectly. It holds on deduplicated, never-seen text, so it is not memorisation; the task is simply easy. Read it as "the pipeline works end to end", nothing more.

2. The priority score is the real one, and it is measured against topic

Priority is where a classifier can actually be useful, and where it can also cheat: VPN tickets skew high-priority, printer tickets skew low, so a model that only recognises the topic already scores 0.513. This one scores 0.901 — the 0.39 on top comes from reading how many people are affected and how soon it matters.

That baseline was not free. An earlier version of the generator sampled urgency and impact but wrote the body from the category alone, so the text carried no severity at all. A model trained on it scored 0.469 against a 0.508 majority baseline — worse than guessing. The fix was in the data, not the model: glpi-testkit now phrases severity the way users do.

3. Train and test overlap unless you force them apart

The generator draws from a finite template set, so unique texts saturate:

Different random seeds are not a split. Measured at 9000/3000, most of the test batch had been seen verbatim in training. train_test_split therefore deduplicates by default — and you can see the inflated numbers yourself with deduplicate=False.

Even deduplicated, a test ticket is a template sibling of a training one: same sentence pattern, different slot values. So this measures generalisation across slot values, not across writing styles.

What that adds up to

These figures will not transfer to your service desk. They establish that the pipeline trains, that it reads severity rather than just topic, and that it does so equally well in three languages. Getting a number that means something for your tickets means fine-tuning on your tickets.

Installation

pip install glpi-ticket-triage

The distribution is glpi-ticket-triage; the command it installs is ticket-triage, and the module to import is ticket_triage. The glpi- prefix keeps it in line with the rest of the suite and avoids a name PyPI rejects as too close to an unrelated project.

Use it

ticket-triage train -n 12000 --report artifacts/metrics.json
ticket-triage evaluate
ticket-triage predict "Impossible de se connecter au VPN, tout le site est bloqué"
ticket-triage predict - --json < ticket.txt
ticket-triage sample -n 5

from ticket_triage import TicketTriageModel, train_test_split, evaluate

train, test = train_test_split(12000, seed=42)
model = TicketTriageModel.train(train.texts, train.categories, train.priorities)

print(evaluate(model, test, train).to_dict())
print(model.predict("الطابعة معطلة في الطابق الثالث"))
model.save("artifacts/model.joblib")

The public demo

pip install -e ".[demo]"
ticket-triage train
python app.py

A Gradio page: paste a ticket in any of the three languages, get a category, a priority and the runners-up. It states its own limitations on the page, because a demo that implies more than it delivers costs more credibility than it earns.

Deploys to a Hugging Face Space on the free CPU tier — the model is 0.5 MB.

Why not a transformer

graph LR
    T["Ticket text<br/><i>fr · ar · ary</i>"]

    subgraph FEAT["Features"]
        CH["char_wb n-grams 2-5<br/><b>survives code-switching</b>"]
        WD["word n-grams 1-2"]
    end

    subgraph CLF["Two classifiers"]
        C1["category<br/>logistic regression"]
        C2["priority<br/>logistic regression"]
    end

    OUT["category + priority<br/>+ confidences"]

    T --> CH & WD
    CH & WD --> C1 & C2
    C1 & C2 --> OUT

    style FEAT fill:#e8f0fe,stroke:#1a73e8,stroke-width:2px
    style CLF fill:#e6f4ea,stroke:#1e8e3e,stroke-width:2px
    style CH fill:#fce8e6,stroke:#d93025,stroke-width:2px

Three constraints decide it:

• No GPU, on whatever server the client already has. This trains in 50 seconds on a laptop CPU and the artifact is 0.5 MB.
• Code-switching. A Moroccan ticket mixes Darija in Arabic script, Darija in Latin letters, French and English — sometimes in one sentence. Word features fragment across that; character n-grams see imprim in imprimante and طابع in الطابعة as ordinary substrings either way. The per-language table above is what that buys.
• The data is synthetic, which bounds what any model can learn from it. Spending a transformer's compute on template text buys accuracy on the templates.

Swapping in a transformer means replacing model.py and nothing else — the CLI, the evaluation and the demo talk to TicketTriageModel only.

Privacy

No customer ticket was used at any point. Every training example comes from glpi-testkit, whose templates are committed in plain sight. There is no source dataset, so there is nothing to leak — which is what makes both the model and this repository publishable.

If you need to clean a real export before working on it, see glpi-anonymizer and presidio-mena.

Contributing

See CONTRIBUTING.md.

git clone https://github.com/novagiosAI/ticket-triage.git
cd ticket-triage
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.