Contact usRequest a demo

Providing agents with suggestions in text chats

Since ChatGPT burst on the scene in late 2022, the applications of generating text using large language models (LLMs) have been explored in considerable depth. Some of the results have been very impressive and provoked a lot of interest. At the same time, the shortcomings of generative models have been widely reported. For example, ChatGPT is prone to provide incorrect responses to even simple questions. Sometimes, its answers bear almost no relation with reality.

If your organization is considering using a generative LLM to assist your clients, you must strike a balance between the potential benefits and pitfalls of employing this technology. One solution is to vet the model’s responses before they’re sent to customers.

Chat suggestions let you do exactly that. Agents can request suggestions for how to respond to a visitor’s chat message. They can then review the suggestion they receive.

  • If they’re happy with the suggestion, they can send it to the chat right away.

  • If it needs editing, they can do so before sending it.

  • If the suggestion isn’t helpful, they can reject it. In that case, no trace of the suggestion remains in the chat.

This approach enables agents to benefit from all the sources of information available to the suggestion bot and can help them answer customer enquiries more quickly. At the same time, it reduces the risk of your bot giving customers and prospects incorrect information.

This is a preview feature. It may be subject to change or removal with no further notice.

To enable preview features, set com.unblu.platform.enablePreview to true.

For more information on preview features, refer to the Unblu release policy.

Suggestion sources

A suggestion source is the endpoint Unblu calls when an agent requests a suggestion.

There are two possible sources for chat suggestions. You can:

You can manage suggestion sources in Unblu through the Account Configuration interface and the SuggestionSources web API.

Besides the usual endpoints to create, read, update, and delete suggestion sources, the web API also has endpoints to search for suggestion sources, to get a suggestion source by name, and to ping the endpoint of a suggestion source.

The same options are available in the Account Configuration interface. For more information, refer to the Suggestion sources section of the Account Configuration interface guide.

Using the UCB for chat suggestions

To use the UCB for chat suggestions, set the URL of your suggestion source to point to the endpoint /unblu/suggestions in the orchestrator. If the orchestrator URL is https://orchestrator.conversational-bot.yourcompany.com, the endpoint for the suggestion source would be https://orchestrator.conversational-bot.yourcompany.com/unblu/suggestions.

You must also set the following environment variables in the orchestrator’s configuration:

  • UNBLU_SUGGESTIONS_ENABLED: Enables the chat suggestion feature on the Collaboration Server and registers itself as a suggestion source.

  • UNBLU_SUGGESTIONS_TIMEOUT: Timeout for chat suggestion requests from the Collaboration Server to the Knowledge Bot.

  • UNBLU_SUGGESTIONS_URL: The URL to the chat suggestion endpoint in the orchestrator, including the port and ending in /unblu/suggestions/.

Integrating your own suggestion bot

If you choose to integrate your own suggestion bot, the endpoint Unblu calls to request suggestions must be capable of processing the ChatSuggestionRequest Unblu sends in the body of the request.

The body of the response your suggestion bot sends to Unblu depends on whether it was able to generate a suggestion:

Enabling suggestions

Once you’ve added one or more suggestion sources to Unblu, you must enable them in the conversation templates of the conversations you want agents to be able to request suggestions in.

See also

  • For information on adding suggestion sources in the Account Configuration interface, refer to the Suggestion sources section of the Account Configuration interface guide.