From what I have seen, all utterances have to be manually entered by us. It would be great if AI could generate additional utterances for us from the examples we give it about any intents we're creating. I've used another chatbot builder and this was a feature it had that made creating a variety of utterances easy as the AI can create ones you wouldn't think of quickly. Have a button to click and automatically generate 5, 10 or 20 additional utterances. This way you know you're covering a wide array of things a user might say. Thanks.

Can you tell me more about you use case? Why don't you use AI Task to detect the intent? or AI transition?

I believe you tagged the wrong person here Bassam.

Nope, I tagged the one who added an emoji 🙂

Ohhh haha! Wow good catch with the Emoji.

😄

Intent is not an AI task. A competiutor sdoes that acutally nicely - genrate the utterances via AI (so the user then can delete them) and also has negative utterances that can be added (i.e. if you THINK it is that intent, but one of those trigger, then no, it is not that intent)

Can you give me an actual example that can't be achieve with AI task or AI Transition?

simple. YOu assume an AI task, which is AN AI run. I have a simple support bot and my Intent switch has right now 8 intents. They are way faster than AI integration and provide less cost. Problem is that for every of the 8 I need to sit down and think of all kinds of wordings. What he asks for is basically that the designer can be asekd to generate another 20 variatns of the same intent, which the user can then filter. I know - and to a degree I agree - that AI is a solution for that - ONE DAY. Right now AI is still slower and costier, and by your words one could well remove the intent at all, instead using a (btw., quite complex to get right) AI task to find which of the X intents it mAY be (something olde models like 3.5 may easily hallucinate in). There is a good reason to keep the NKP taks aeound - mostly final user say, like taking the 20 generated utterances and eliminate bad ones.

I feel like I may be in a corner here

Just saw the new AI based intent determination and seriously, whoever thought that out never did consider limitations in complex cases (which seems to be standard for those things). YOu literally have to rely on the AI to take the meaning from a short intent - there is no way to even have alterantive utterances without cluttering your workflow. And GPT 3.5 is known problematic for certain things - 4.0 ios better at it - that are too simple. Like the auto translation (which seems to change languages on email addresses and other short sentences due to not considering to filter non-sentences out) - I would consider it close to unusable in anything except the most obvious edge case.

How would you use AI Task or Transition to detect intent? See the response below from @cool-policeman-94412. He seems to know which one I'm referring to. Also, the negative intents are a good one to note. When creating intents in Botpress we have to enter a list of utterances so instead of manually entering a list of utterances, AI could generate a range of utterances based on the few-shot examples we give it. In the screenshot, I create an intent but I have to enter ALL the phrases I think the user will use. Rather than me having to enter all that I think, AI would be able to generate a list of possible utterances based on the few examples I've given.

Simply speaking - go to voiceflow, see how they do it.

Lol, that's the one.

I agree with you theoritically, but can I have real examples? I just want to try on my end. As I said I want to learn that's all 😉

Hey, do you still need examples or did the screenshot I provided suffice? In the screenshot, it says, "add 7 more utterances to make" it more resilient. So, the idea is, instead of the bot builder needing to** 'think' **of 7 more utterances a user could make, there is an AI function enabled where I can click a button to generate 7 or more utterances based on the examples I've already given it. In this case, I gave it three. Hopefully, that makes more sense. And then the negative intents as you've noted.

@clean-xylophone-13519 I understood your request and I agree to it

I don't have a real case scenario where the AI Task doesn't work but wouldn't it be the AI Transition I'd use to detect intent? Now, the question I have coming up is when do we use intents vs. AI Transition which also detects intent? Perhaps @bumpy-butcher-41910 can shed some light on this. If there is a tutorial on this, please direct me to it. Thanks

I had a problem with different appointments depending on wording. This is where the AI limitations show because you have not enough control to tell it what is NOT a particular intent.

Sorry I didn't get that

It is what I said - we have different flows based on intent, and it was getting confused between an intiial call and a service request for existing customers at this point. Could not really rule it out becuase i lacked negative intent to limit them - too similar otherwise.

kk, so it was about negative intent too. Understood. thanks

Because AI is not human. it still makes way mre errors and sometimes intents are quite similar.

I understood that, I still can't think of business need. I just want an example to imagine it

We have multiple types of appointment in a bot. Sadly the AI seems to ignore the various ways people address them, often falling back to the first.

We are going in circles 😄

We are not going in circles. YOu ask for details that are non public. SImple like that. Maybe read the voiceflow documentation on their implementation? Here is a link that also gives a use case: https://learn.voiceflow.com/hc/en-us/articles/18749046327821-The-NLU-Fallback-Intent-Avoiding-False-Positives