The following shows an example of part of an entity extraction based call flow:

Here is what it is doing:

• Introducing the process using some Realtime TTS 
• Asks the caller to input their customer number by speech
  • If the caller speaks his customer number, then check if it is found
  • If not, we ask for the customer number again, applying the same logic

Let's look at the Input ASR Object we are using:

This will:

• Say to the customer "What is your customer number" using the TTS
• Wait for input from the customer
  • For a maximum of 5 seconds, if the customer says nothing
  • If they start to speak, then for a maximum of 20 seconds
  • Once they have started to speak, then a pause of 2 seconds will cause the ASR to deliver a "final" result containing the digits we require

The second Input ASR Object is very similar, just this time we prompt the user letting them know we didn't understand what they said:

Following this, we search for the customer number in the database (here we use a list, but this could be anything - a REST API, DB Call or whatever is needed.

This is generally a good pattern for such call flows:

• Ask the caller for the data you require
• If you do not get it, ask one more time
• Chat the data you get for validity

Key points:

• You can cascade this approach, asking for several things one after the other.
• If you are combining this approach with an LLM, simply add a check whether the appropriate entity is already filled with data, and skip that part of the call flow if you already have the data.