Quality: test suites and guardrails

Before promising a behavior to your customers, replay it. A test suite is a list of sample conversations with the expected outcome; each run really calls the agent's model and tools, then produces a report card.

• Ready-made suites: banking, telecom, e-commerce, insurance, healthcare, public services. Importing one on a bot flags missing or unconfigured tools.
• Replay on demand and name the iteration: the "Quality proof" page tracks progress from one report card to the next.
• Some suites come with a separate validation set: do not use it to tune the prompt, it is there to check the improvement holds.
• Per-agent guardrails: masking of personal data in logs (maskPiiInLogs) and prompt-injection detection in three modes — off, log or block.
• A run uses calls to your AI provider: launch it after every change of prompt, tool or model.

TOKEN="<accessToken>"   # voir « Référence API »

# Importer le jeu de tests « banque » sur un bot (clés : GET /admin/evals/builtin)
curl -X POST https://api.votredomaine.com/api/v1/admin/evals/bot/VOTRE_BOT_ID/builtin/bank-advisor \
  -H "Authorization: Bearer $TOKEN"

# Lancer un rejeu et nommer l'itération
curl -X POST https://api.votredomaine.com/api/v1/admin/evals/suite/SUITE_ID/run \
  -H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" \
  -d '{ "label": "après correction du prompt" }'

# Historique des bulletins
curl https://api.votredomaine.com/api/v1/admin/evals/suite/SUITE_ID/history -H "Authorization: Bearer $TOKEN"