E-commerce

Checkout help is where retail AI meets money. A shopper asks why a card was declined or what happened to an order, and the answer needs real account context. This starter is the guardrail for that context. It masks payment token strings, refuses three named phrases outright and scores every answer before delivery.

A shopper-facing assistant answers anyone with a browser. The people who reach for this starter run search help, cart questions and checkout support at retail volume. Some of that traffic will probe the assistant for the customer table or for a stored code. The configuration assumes that mix instead of a careful internal pilot.

Screening for injected instructions happens before the model is called. That order matters, because a crafted instruction cannot talk its way past a check that already ran. Next the redaction step masks anything that looks like a payment or card token, so the raw value never leaves the gateway. Masking keeps a working answer for the shopper.

Three phrases are blocked outright rather than masked: a private key, a bulk customer export and a gift card pin. A request that contains one of them is refused, and the shopper sees a refusal instead of an answer. The list is aimed at two abuse shapes. One is the attempt to turn a support assistant into a data export tool. The other is the attempt to read stored value codes out of an order record.

Keyword safety screening ships with no custom terms, so the built-in list applies and nothing more. It compares words and phrases only, with no classifier behind it. On the way back, output scoring holds each answer to a combined score of 0.78, with relevance at 0.78 and accuracy at 0.80. The two metrics count equally. An answer under 120 characters or a single sentence fails the same gate and is blocked.

Audit marking closes the sequence and never blocks a request. It notes that the checks ran, and storage plus retention remain settings your team owns. The starter adds no role check, so caller identity rules stay your decision. Card handling and fraud scoring sit outside this starter and must keep their own controls. Choose the Marketplace starter when buyers and sellers need separate treatment, or Retail Customer Service when agents draft the replies.