Trier les tickets avec Jev, appliquer ses règles en Python et laisser le LLM

The following notes reconstruct a practical path around “Trier les tickets avec Jev, appliquer ses règles en Python et laisser le LLM préparer la réponse : un tutoriel concret avec NOVA.”. Emphasis stays on contracts, checks, and drop-in code placeholders rather than motivational framing. When working through the Overview stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion.

Jev ne remplace pas le modèle qui rédige

The Jev ne remplace pas stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Pin the interpreter and dependency lockfile before teaching the loop. Drift between laptop and CI is the most common silent break for API demos.

L’architecture avant le code

The L architecture avant le stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Pin the interpreter and dependency lockfile before teaching the loop. Drift between laptop and CI is the most common silent break for API demos.

Message client + historique
            ↓
Jev : service, urgence, frustration
            ↓
Python : validation et règles métier
            ↓
Agent LangChain + LLM
            ↓
Consultation des commandes et procédures
            ↓
Dossier pour l’équipe support + brouillon de réponse

Préparer l’environnement

The Pr parer l environnement stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Pin the interpreter and dependency lockfile before teaching the loop. Drift between laptop and CI is the most common silent break for API demos. The Pr parer l environnement stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion.

TYPESAFE_API_KEY=ta_cle_typesafe
OPENAI_API_KEY=ta_cle_openai
JEV_MODEL=jev-latest
OPENAI_MODEL=gpt-4.1-mini

jev = TypeSafeClassifier(
    model=os.getenv("JEV_MODEL") or "jev-latest",
    timeout=30,
)

modele = ChatOpenAI(
    model=os.getenv("OPENAI_MODEL") or "gpt-4.1-mini",
    timeout=60,
    max_retries=1,
)

Les trois primitives : Choice, Noul et Score

For the Les trois primitives Choice stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Separate client construction from the message loop so providers can be swapped without rewriting the conversation state machine.

Choice : choisir un service

For the Choice choisir un service stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Separate client construction from the message loop so providers can be swapped without rewriting the conversation state machine.

Noul : évaluer une question oui/non

For the Noul valuer une question stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Separate client construction from the message loop so providers can be swapped without rewriting the conversation state machine. For the Noul valuer une question stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion.

Score : situer la frustration sur une échelle

When working through the Score situer la frustration stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Log request id, model id, and latency on every call. Without that trail, intermittent provider errors look like application bugs.

from langchain_typesafe import Choice, Noul, Score

def questions_triage():
    return {
        "service": Choice(
            instructions=(
                "Quel service doit traiter en priorité la dernière demande du client ? "
                "Utilise le contexte seulement pour comprendre cette demande."
            ),
            criteria={
                "livraison": "Retard, suivi ou réception d'une commande.",
                "facturation": "Paiement, facture ou remboursement.",
                "technique": "Panne ou utilisation d'un produit.",
                "autre": "Demande ambiguë ou sans rapport avec les catégories précédentes.",
            },
        ),
        "urgence": Noul(
            instructions=(
                "Les faits décrits nécessitent-ils une prise en charge immédiate, "
                "plutôt qu'un traitement normal ? Ne te fonde pas seulement sur le ton."
            )
        ),
        "frustration": Score(
            instructions="Quel niveau de frustration le client exprime-t-il ?",
            criteria=[
                "Le client s'exprime calmement, sans insatisfaction.",
                "Le client exprime une insatisfaction tout en restant mesuré.",
                "Le client exprime une forte colère ou des réclamations répétées.",
            ],
        ),
    }

Analyser un premier ticket

When working through the Analyser un premier ticket stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Log request id, model id, and latency on every call. Without that trail, intermittent provider errors look like application bugs.

reponse_jev = jev.invoke(requete_triage(ticket))

print("Service :", reponse_jev.choices["service"].choice)
print("Urgence :", reponse_jev.nouls["urgence"].noul)
print("Frustration sur 2 :", reponse_jev.scores["frustration"].score)

Service : livraison
Urgence : 0.76
Frustration sur 2 : 1.93

Les règles métier restent dans le programme

When working through the Les r gles m stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Log request id, model id, and latency on every call. Without that trail, intermittent provider errors look like application bugs. When working through the Les r gles m stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion.

def orienter_ticket(analyse, seuil_urgence=0.8, seuil_confiance=0.6):
    raisons = []

    if analyse["urgence"] >= seuil_urgence:
        raisons.append("urgence élevée")
    if analyse["frustration"] >= 1.5:
        raisons.append("forte frustration")
    if analyse["confiance_service"] < seuil_confiance:
        raisons.append("service incertain")
    if analyse["service"] == "autre":
        raisons.append("demande à clarifier")    return {
        "service": analyse["service"],
        "priorite": "haute" if analyse["urgence"] >= seuil_urgence else "normale",
        "revue_humaine": bool(raisons),
        "raisons": raisons or ["traitement courant"],
    }

{
  "service": "livraison",
  "priorite": "normale",
  "revue_humaine": true,
  "raisons": ["forte frustration"]
}

Donner à NOVA des informations à consulter

The Donner NOVA des informations stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Pin the interpreter and dependency lockfile before teaching the loop. Drift between laptop and CI is the most common silent break for API demos.

Assembler l’agent LangChain

The Assembler l agent LangChain stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Pin the interpreter and dependency lockfile before teaching the loop. Drift between laptop and CI is the most common silent break for API demos.

def creer_agent(modele, analyse, orientation):
    contexte = json.dumps(
        {"analyse_jev": analyse, "orientation": orientation},
        ensure_ascii=False,
    )

    return create_agent(
        model=modele,
        tools=[consulter_commande, consulter_procedure],
        system_prompt=(
            ROLE_NOVA
            + "\nContexte de traitement fourni par le programme :\n"
            + contexte
        ),
    )

Ce que montrent les tests du notebook

The Ce que montrent les stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Pin the interpreter and dependency lockfile before teaching the loop. Drift between laptop and CI is the most common silent break for API demos. The Ce que montrent les stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion.

suivi = traiter_ticket(
    "Quel article contient cette commande ?",
    modele,
    jev,
    historique=dossier["messages"],
)

print(suivi["reponse"])

Ce que je garderais pour un vrai projet

For the Ce que je garderais stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Separate client construction from the message loop so providers can be swapped without rewriting the conversation state machine.

Operational checklist

For the Operational checklist stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state.

Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline.

Separate client construction from the message loop so providers can be swapped without rewriting the conversation state machine.

Cache stable system instructions and tool schemas. Re-sending identical preamble is a common source of burn.

Pin dependency versions and record the image digest that ran the demo. Reproducibility beats tribal knowledge.

Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph.

Before promoting the stack, freeze versions, capture a golden transcript for the critical path, and confirm rollback steps. Shared environments need rate limits, tenancy checks, and a clear owner for secret rotation. Prefer boring reliability over clever one-off demos.

Batch note for cf51dd985f5e: keep provider keys out of the repo, set a per-session token ceiling, and store transcripts next to the eval fixtures so later model swaps stay comparable.