01

Test source-grounded product answers

Ask questions whose answers can be checked directly against product pages, variant data, delivery information, return rules, or other connected sources.

02

Include variants, policies, and compatibility

A useful evaluation includes size, color, availability, materials, compatibility, delivery, returns, and comparison questions rather than only simple product titles.

03

Check how uncertainty is handled

When a source is missing or ambiguous, the assistant should ask for clarification, state the limitation, or use the approved human route instead of guessing.

04

Review product cards and links

Check that product cards, destination links, option information, and availability context match the current storefront experience.

05

Verify the operator workflow

If Telegram operator notifications are configured, confirm what context reaches the operator and remember that a human decides manually whether and how to respond.

06

Use a repeatable evaluation set

Keep a representative question set and rerun it after catalog, policy, or configuration changes. Do not treat one demo answer as a universal accuracy guarantee.