Chat agent

Testing and publishing

Test the agent in the Test Chat before customers see it, publish your changes when you're confident, and repeat the check questions after every major change.

View as Markdown
Video: Test the agent and fix its answers AC-08Testing the draft in the Test Chat, publishing, fixing an answer with an FAQ.

#Draft and published version

In the Playground you work on a draft. Customers talk to the published version, which changes only when you press Publish: saving is not enough.

In the Playground, the draft status at the top right of the canvas tells you where you stand: “Unsaved changes”, “Saved” or “Published”.

Only the Playground gets published. Other changes take effect at a different moment:

What you changeWhen customers see it
Instructions, scope, model, tools and nodes (Playground)After Save changes and Publish
Training: website, documents, FAQs, textsWhen the training is “Completed”
Agent profile: interface, limits, live chat, contact formAfter Save changes in the tab

#Testing with the Test Chat

The Test Chat is the agent's chat, opened inside the Playground. It lets you test the saved draft before you publish it: customers see nothing until you publish. Changes you haven't saved can't be tested.

  1. Save your changes

    In the Playground, press Save changes.

  2. Open the Test Chat

    Press Test Chat at the top, from any page of the agent. The Playground opens with the Test Chat on the right; at the top of the panel you read “You are testing the saved draft of the agent.” If the draft is the same as the published version, or if Live is on, you read “You are testing the published version, the one that answers your customers.” instead.

  3. Write like a customer

    Ask the test questions. If you added tools, the Playground's extra capabilities, also make the requests that should trigger them.

  4. Fix and try again

    Change the instructions or the nodes, save and write again: the conversation continues with the updated draft. To start over, choose New chat in the chat menu. Close the panel with the X.

When you're satisfied, publish. If the draft has errors, the errors panel opens instead of the chat: fix them and try again.

Officina Rossi's Playground with the canvas on the left and the Playground Test Chatbot panel on the right: the note “You are testing the saved draft of the agent.”, a question about the price of a car service and the agent's answer.
The Test Chat tests the saved draft before it is published.

#What the Test Chat looks like

It looks like the chat on your website: colors, image, welcome, suggested questions, attachments and voice messages, if you turned them on. It doesn't show the privacy consent request or the initial contact form. The chat menu has New chat. It also has Request an operator, if in Operator Chat you turned on the live chat and its item in the chat menu.

#Where test conversations end up

  • In Chat agentConversations, when you choose the “Internal” channel. The normal lists don't show them.
  • They don't count in the Dashboard statistics and don't send notification emails.
  • The agent's replies use up the plan's AI messages, like replies to customers.
  • An operator request stays waiting: no emails or notification sounds go out.
  • A submitted Form really creates or updates the contact in the contact book: test with made-up data, such as Mario Rossi.

Tests trigger automations. For the workspace's automations, a test counts as a customer's chat. They are triggered by the agent's messages, tools, Forms, operator requests and new contacts. If an automation sends emails or messages or makes calls, they really go out, and the execution counts toward the month's quota. Before testing, check which automations are active in the Automations panel of the Overview.

#Publishing your changes

Before you start

  1. Save the draft

    Press Save changes: “Agent saved” appears and the status becomes “Saved”.

  2. Open the publish window

    Press Publish: “Publish Agent” opens.

  3. Confirm

    Press Confirm and Publish.

Result

“Agent published” appears and the status becomes “Published”. On all channels, new conversations and those already open use the new version from the next message.

To look at the published version without touching the draft, use the Live switch: see Live mode.

The chat link shows you the agent as a customer sees it: with the published version, the privacy consent and the contact form. Use it after publishing, as a final check.

  1. Check that the agent is active

    The status button at the top must say “Active”: while the agent is paused, the link doesn't work. See Active and Paused.

  2. Copy the link

    In Chat agentAgent profileShare, “Link and Direct Sharing” panel, press Copy.

  3. Open it in a private window

    Paste the link into a private (incognito) browser window: the chat treats you as a new customer.

  4. Ask the test questions

    Then read the conversation again in Chat agentConversations.

Chats opened from the link are real conversations: they appear in Conversations and in the Dashboard like those from your website, and they use up the plan's messages. You can delete them after testing: see Exporting and deleting.

#The test questions

Before opening the agent to customers, test it with 20 real questions: the list with examples is in The test with 20 real questions.

Ask the same questions again after every major change. If you added tools, add the requests that should trigger them.

#When an answer is wrong

Facts are fixed in the knowledge base, behavior in the instructions and tools. How to find out where an answer comes from, the typical cases and the quick fix with Correct answer are in Why the agent answered that way. Below are the cases you fix in the Playground.

What you noticeWhere you fix it
Made-up informationIn the instructions: ask it to use only the knowledge base. Execution must be set to Knowledge Base.
Off-topic answersIn the instructions, under “What not to do”, and in the business scope.
It doesn't remember what the customer said earlierNumber of messages must not be None.

After the fix, ask the same question again.

#See also