How an answer is produced

A visitor sends a widget message. When AI is enabled and the conversation is in the relevant mode, the server sends a request to the selected provider: instructions, knowledge-base text, the visitor’s message and recent conversation history. The response is saved in the conversation your team can view.

Choose an available provider

Options include server-configured GoMsg presets and your own compatible API. Server presets depend on an administrator configuring credentials and models. Your own API needs an endpoint, key and supported model name; verify compatibility with a real request.

  • Check your server and provider terms, including request limits.
  • Keep the API key out of knowledge-base text, greetings and website code.
  • Fast and Smart name configuration options; they do not guarantee response time or quality.

Configure the widget step by step

  1. Sign in to GoMsg and open your site in the widgets section. You need widget management permissions and access to the section on your server.
  2. Open AI settings, enable the assistant and choose a server preset or your own API. For a custom connection, enter the endpoint, key and model.
  3. Set the assistant’s name and instructions: what it helps with, its tone, and questions it should pass to a person.
  4. Add current knowledge-base text: services, prices, hours and policies. Organise facts into short, clear items.
  5. Set the reply limit, handover keywords and call-operator button. Enable the AI message badge if appropriate.
  6. Save and send several test questions through the published widget. An appearance preview does not test model replies.

A knowledge base helps; it does not eliminate mistakes

A model may misunderstand a question, miss a condition or invent an answer. “Do not invent facts” is a useful instruction, not a guarantee. Test prices, deadlines, exceptions, incomplete questions and ambiguous wording.

  • Ask a question not covered by the knowledge base: the assistant should offer a staff member’s help.
  • Test each language you need; quality depends on the chosen model.
  • Repeat tests after changing models, instructions, services or prices.

Human handover and reply modes

Settings include handover by reply count, keywords and visitor request. A staff member can take the conversation and continue with its history. Requesting an operator does not mean a person has already joined; your team needs a response process.

  • Test the call-operator button and selected keywords.
  • Do not treat Human first or Fallback as a guarantee of staff-presence checks or a configured waiting period in the current implementation.
  • Explain normal response hours in the greeting and test when the team is unavailable.

If AI does not reply

An invalid key, unsupported model, provider limit or connection failure can prevent a reply. This does not mean the visitor’s message was not sent: check GoMsg, reply manually and ask your administrator to inspect configuration and error logs. Keep secret keys out of error reports.

What the model receives

The selected provider receives the question, knowledge-base text, instructions and recent conversation context. Avoid passwords, keys and unnecessary personal data. Decide which processing conditions suit your business before launch; self-hosting alone does not make an external model local.