V1.4 Beta Build 38

[gbtinc]

Hello,

We upgraded to the latest AI update and the simplification of the agent instructions broke down our original agent instructions. The agent kept having trouble transferring calls.

After hours to modifying our agent instructions to get the AI to work again, we realized that 3cx may have hard coded the following:
15:14:58.228 HIGHLIGHT No progress after 60s - no meaningful tool called. Sending warning.
15:15:28.229 HIGHLIGHT No progress after 90s - dropping call.

I cannot seem to find anywhere to update this in the 3cx configuration. The agents were already set to a 10 minute call duration and we even changed it to 15 minutes just to overwrite.

This is a problem for us because we use the AI to handle basic questions first before handing off to our call center. This helps alleviate FAQs to our call center. 60s and 90s is way too short for a meaningful conversation.

Please advise.

I will be reverting back to v1.2 Build 5 in the mean time.

Thanks.

 

[nikolascx]

Hi yes this is a problem.
But from where is your Agent answering questions? Surely not from the vector storage. Otherwise this would not happen.
Can you explain the flow?
And what questions is the AI Answering? Questions like "What is the weather?"

What is happening here is that the AI classified your call as a call where "no Meaningful progress is being made".
It's the same as calling someone and staying silent gibberish.

If the basic questions were coming from a knowledge base, this problem would not occur.

Can you put here a flow to understand what questions are being asked. Thanks so much

 

[gbtinc]

the questions asked to the agent are knowledge source questions.

the flow is DID > AI Queue > AI Agent. AI agent answers questions about customer service and technical support first unless the caller specifically requested a transfer to those department queues.

## Resolve Before Routing
When a caller states a question, issue, or need that can be addressed with general product, service, company, or policy information, attempt to resolve it directly using your available knowledge sources before initiating any transfer. A transfer is a last resort, not a first response.
After providing an answer, ask:
> "Does that answer your question, or would you like me to connect you with someone who can assist further?"
If the caller is satisfied, close the call courteously. If they want to be connected, proceed to Call Routing.
If the caller responds with detail showing the basic guidance has not resolved their issue (e.g., they have already tried the steps you suggested, or their situation is more specific than your knowledge can address), do not press for more information and do not threaten to end the call. Acknowledge what they have tried and offer to connect them with the appropriate department.
Only proceed directly to routing without attempting to resolve when:
- The caller explicitly asks to be transferred, connected, or to speak with a person, department, or extension.
- The caller's reason matches a department that exists in the address book (route immediately to that department).
- The caller's question is entirely outside the scope of your knowledge sources.

I can DM you some call samples or transcriptions.

 

[nikolascx]

But does the AI Answer from the knowledge source?
I think it interprets knowledge source by "it's training knowledge"

## Resolve Before Routing
When a caller states a question, issue, or need that can be addressed with general product, service, company, or policy information, attempt to resolve it directly using your available knowledge sources before initiating any transfer. A transfer is a last resort, not a first response.

Change it to

When a caller states a question, issue, or need, use vector_store_search before initiating any transfer. A transfer is a last resort, not a first response.

If you still fail get me the log.

 

[gbtinc]

the AI does answer properly from the knowledge source.
At around 60 seconds, even if the AI is in the middle of repeating the knowledge source, it'll say: I'm having a bit of a hard time understanding exactly what you need. at 90 seconds, it hangs up.

After switching to vector_store_search, it seems to fix the issue. will continue to test for now and hold off on reverting the backup.

Another thing is v1.4b38 seems to have issues transferring to extension numbers and names. Previously on this ticket, https://www.3cx.com/community/threa...s-to-transfer-to-can-transfer-by-name.137135/, we were able to address the issue by adding extensions to the call routing. This no longer seems to work on v1.4b38. It sometimes cannot find an extension and/or the name pronunciations are no longer as accurate as before. On the previous version, when the caller ask to be transfer to a name, the AI did not have issues with uncommon names. On this new version, it sounds like it has a hard time understanding or crossreferencing the name.

 

[nikolascx]

I saw your log. Thanks However I understand your concern. And I can explain.
We have a receptionist, and PA at the moment. Customer Service Agent is coming soon.

The customer service agent of course will try to solve issues before trying to transfer, but the Receptionist and the PA are designed to route calls.

For now, put some sources to help your customers solve issues. Even if it is a text file with a bunch of top 50 solutions. like this you can convert the receptionist into a support agent.

The AI Is answering from "ITs knowledge source" - from what it was trained on, not from your company knowledge source.

About the second comment

Another thing is v1.4b38 seems to have issues transferring to extension numbers and names. Previously on this ticket, https://www.3cx.com/community/threa...s-to-transfer-to-can-transfer-by-name.137135/, we were able to address the issue by adding extensions to the call routing. This no longer seems to work on v1.4b38. It sometimes cannot find an extension and/or the name pronunciations are no longer as accurate as before. On the previous version, when the caller ask to be transfer to a name, the AI did not have issues with uncommon names. On this new version, it sounds like it has a hard time understanding or crossreferencing the name.

Yes we are working on it. Coming very soon.
For now, help it by adding topics.. Also you have an error in the log related to transfers which I am checking. For this send me your prompt so I can understand why it is failing.

I suggest you use the prompts we ship. And build on that. Be careful where you place things. I will make a video on how to build this shortly.

 

[gbtinc]

the prompt we created is a cross between old and new. i will DM it to you.

 

[gbtinc]

i ended up deciding to roll back to a complete backup that i took within 3cx prior to upgrading to AI 1.4b38, however, after the restore completed, it still seems to have AI 1.4b38 installed. Any ideas why?

I had to do a restore of the virtual machine instead.

 

[nikolascx]

But the backup does not remove the installed binaries.
If an update is installed, the restore will not help.
Would have been better to put a markdown file and put answers in there.