LLM Intents

LLM stands for Large Language Model. The DialoX platform supports all major LLM providers under the hood. By default the platform uses Openai chatGPT on Azure and Google Gemini, both hosted on European servers.

Add or edit intents

You can add a new intent by clicking on the Add intent button. Now you can specify the label and description of the intent.

The new LLM intents rely far more on the description of the intent than on the training phrases. It is important to give a clear instruction on when to match on the provided intent, as if you were instructing another person.

For instance, if you'd like to create an intent for switching language mid-conversation you can provide a description such as:

"When the user asks to switch to a different language or wants to speak in a different language."

Descriptions don't necessarily need to be translated to all the languages your Smartapp supports. Make sure to offer at least a text in the first language of the bot. The LLM will know how to interpret this description for all the other languages out of the box.

It's no longer necessary to provide extra examples, although they're still provided to the LLM when prompting it to match on intents.

Enable 'only use locally' when you are using the intent only to trigger the next link in an Ask or Prompt node.

Disable 'only use locally' when you are using the intent to trigger outside an ASK or PROMPT node. These only make sense for usage in Intent triggers. Because these intents can be triggered without knowing the exact context upfront.

Local intents

Toggle Only use locally when you need to recognise a user intent within the context of a specific dialog or flow. If you want this intent to be able to be triggered anywhere in your bot you need to keep it global (so toggled off).

When you make sure the description is clear, you probably don't need to add any examples. You can try testing your intent in the Intent tester pane on the right hand side.

Testing intents

Intent testing can be quite a tedious process. The AI will use its algorithm to calculate which intent to classify for any given user utterance. The way that is done by the AI remains a black box. This often requires quite some tweaking of the description before the required behaviour is achieved.

It is possible to test the classification of intents by running your bot and see if it behaves as expected. Another, faster way is to test the intent classification via the Intent tester in the right side panel.

By typing a sentence (user utterance) in the Sentence field in the top of the side panel on the right, the system will automatically run it against the AI.

In the above example we are testing the global intent for opening hours.

For local intents, you need to specify the context by adding the intents you want to test the utterance against in the Context field below.

Difference with legacy Q&A intents

The previous Q&A intents required you to provide a set of example sentences to match the correct intent. However as there are countless ways to ask the same question, it became pretty difficult to provide an example for each and every way something can be expressed.

New LLM intents don't require all these examples, nor does it need any training. You can simply state what is intented by the user, instead of how. This allows for a more brief description and easier configuration of intents.