Try it and releases

Draft and publishedCopy Anchor Link

The AI's configuration has three parts: the knowledge base, skills, and policy numbers (政策数字). Every change you make in the console goes into the draft first.

• Try it (试聊) always uses the draft.
• Before your first publish, customers get the draft too, and changes are live immediately.
• After your first publish, customers only get the latest published version. Draft changes reach customers when you publish again, and you can roll back in one click.

Autonomy levels (AI 自主程度) and working hours are not part of a release; they take effect as soon as you change them in Settings (设置).

Try itCopy Anchor Link

Try it (试聊) is a tab under AI support → Try & release (试聊与发布) in the sidebar, visible to owners and admins.

• It embeds the real chat window, exactly as customers see it, running on your current draft: knowledge base, skills and policy numbers, including changes you have not published.
• The AI's answers are always shown, whatever the autonomy level, so you can see how it would answer a question type that is still at R1.
• Where the conversation would go to a person, a line says so (in Chinese, starting "（试聊）到这里会转给人工客服"); nobody is actually notified.
• Try it conversations do not appear in the inbox, analytics or promotion evidence, and do not count toward your plan's AI conversations. What the model calls cost does count toward the monthly AI spend cap.
• "Chat as" (以谁的身份) lets you chat as an anonymous visitor (未登录访客) or as one of the workspace's customers, to try order questions (the AI reveals no orders to anonymous visitors). New conversation (新对话) starts over.
• "Things to ask" (可以这样问) lists sample questions; Copy (复制) one and paste it into the chat.
• If the AI offers a refund in Try it, you can confirm it as the customer, but your refund endpoint is never called.

PublishingCopy Anchor Link

Releases (发布) is a tab under AI support → Try & release (试聊与发布) in the sidebar. Only owners and admins can publish, force a publish or roll back.

What a release contains: the draft's skills and policy numbers (including each market's own), and every knowledge base entry whose status is Ready (entries still indexing are left out).

After you click Publish to customers (发布给顾客), optionally with a note on what changed:

1. The draft is frozen and the knowledge base is copied for this version.
2. purser waits until the copied knowledge can be searched.
3. Replay evaluation: up to 20 recent, distinct customer questions the AI answered (or tried to) are answered again with the new version, each going through the same check as a real reply. Nothing is sent to anyone.
4. Golden questions are run (Starter and up; see below).
5. The version is published, or blocked.

This usually takes 2–5 minutes; you can leave the page. Only one version can be evaluated at a time. A version's status is one of Evaluating (评测中), Blocked by evaluation (被评测挡住), Published (已发布) or Failed (失败).

What blocks a publishCopy Anchor Link

Compared with the version customers use now:

A first publish has nothing to compare with: it is blocked if there are at least 5 replayed questions and fewer than 50% are answered, or if 2 or more answers are unsafe.

A blocked version lists the questions it got wrong. Look at them in Try it first. If you are sure it is fine, you can Force publish (强制发布) with a reason of at least 10 characters; the reason is archived with the version, which is marked as force-published.

Rolling backCopy Anchor Link

Version history (历史版本) lists the latest 30 versions. Click Roll back to this version (回滚到这个版本) on any earlier published version and customers are pointed back to it at once: it applies from the next conversation, and conversations already running keep the version they started with. Publishing works the same way, so a conversation never changes configuration midway.

Golden questionsCopy Anchor Link

Available on Starter and up. Keep them in the Golden questions (黄金问题) card on the Releases page: questions that must be answered right, each with the right answer. Up to 100.

• Generate from your Q&A: one click turns up to 10 Q&A entries in the knowledge base into golden questions (the Q&A answer is the right answer), never adding a question that is already there.
• The right answer can be longer than a reply, such as a whole FAQ answer: the judge checks that the reply gives the answer to the question (numbers, conditions and outcomes must match); background the question does not ask about is not required.
• A golden question passes only when all of these hold: the reply would pass the check and be sent; the judging model finds it gives the right answer; and it is written in the same script as the question (an English question answered in Chinese fails; languages that share the Latin alphabet are not told apart).
• Results show as correct (答对), not answered (没答出) or missing points (要点不全), with the points and groundedness scores.
• Run on draft (对草稿跑一遍) runs them against the current draft at any time, so you can see whether a change broke something that used to be right.
• Every publish runs them again. A question the previous version answered right and the new one does not blocks the publish (you can still force it with a reason).
• After a downgrade to Free, your golden questions are kept but no longer used to gate a publish.
• Only owners and admins can add, edit and run them.

Start with the questions customers ask most and that cost the most when wrong: return windows, shipping fees, warranty.

Multi-turn simulationsCopy Anchor Link

Golden questions check one question at a time. Real customers ask a follow-up, change the subject or ask for a person. A simulation checks a whole short conversation: an AI plays a customer from your scenario, the same simulated customer used in practice in the learning centre, and chats with the version being tested.

You keep them in the Multi-turn simulations (多轮模拟) card on the Releases page. The plans and permissions are the same as for golden questions. There are up to 20 scenarios, and each has:

A scenario passes when all of these hold:

• the conversation ends the way you expected: answered and the customer is satisfied, or handed to a person;
• the answers contain your points;
• nothing unsafe was said (an overpromise, a contradiction of the sources, leaked information);
• every answer is in the customer's language.

A reply the check would stop counts as a hand-over, because that is what the customer would see. Results show why a scenario failed, and you can open the whole simulated conversation.

• Run (with golden questions) (跑一遍（含黄金问题）) runs the golden questions and the simulations against the draft.
• Every publish runs them again. A scenario the previous version passed and the new one fails blocks the publish.
• Templates give you a start: a return question with a follow-up, an angry customer who asks for a person, and a shipping question in Japanese.

SkillsCopy Anchor Link

Skills (技能) is a tab under AI support → AI abilities (AI 能力) in the sidebar. A skill is an instruction you write for the AI: in a certain situation, what to say, which steps to follow, and where to stop. For example gift wrapping, wholesale enquiries, or pre-orders. Only owners and admins can edit skills.

Each skill has these fields:

• Up to 50 skills per workspace. For each reply the AI sees at most 10 skills' "When to use" lines: first those tied to the current question type, then those not limited to any type.
• Skill text the AI read counts as one of its sources and is checked along with them.
• A skill can only tell the AI how to talk and which steps to follow. It cannot make the AI carry out actions or promise refunds or discounts, and its replies are still subject to the autonomy levels and the reply check.
• Skill changes go into the draft, like knowledge: Try it uses them at once, customers after you publish.