Testing Znuny LLM Bug?

[finalbeta2]

Hi,

I'm testing the new (cool) Znuny LLM add on.
I'm on 7.3.2 for Znuny and 7.3.1 for the Add-on.

I have a couple of issues/questions. And wanted to discus them here, as they may be issues on my side.

Basic Ollama config is working.
1.
When selecting "'FAQ matching & knowledge drafts". This seems to apply in the LLM wizard.
When going back through the wizard. "FAQ matching & knowledge drafts" is deselected again. (Can't see anything in the system logs).

2.
After enabling/disabling Ticket summaries, the ticket summaries keep running on all tickets. When going through the wizard, I can see it's disabled though. But on the tickets, it's still updating the summaries.
At the moment, I would like to be able to update ticket titles, but not create summaries (as this would take up to much of the system).

[Johannes]

Hi,

I will look into this. a 7.3.2 is on the horizion, so this can be fixed there.

Regards

Johannes

[finalbeta2]

@Jonannes, update after testing 7.2.3.

1. = Fixed.
2. = No changes, we needed to manually disable Ticket::Frontend::AgentTicketZoom###Widgets###0000-LLMSummary to stop summeries.
3. = We never got classifications to work. Might be dependend on the full summiers (not stated anywhere). So we reenabled them.
Some tickets show an empty classification block. No logs seem to exist on how to debug. No controls seems to be given on "do we want the AI to classify the ticket, do we really want the AI moving tickets in the system. " etc.
4. = Vectorisation is failing on a number of FAQ's. Somethimes because the plugin should probably set chunking on the request.
5. = Feature request = Filtering where the AI does it's job (mail adressen or queue's) is probably a good option for higher volume systems. This will save a lot of tokens on high volume systems.

If our input is welcome, I'm willing to take this to github? (can we make github issues for the plugins?)

[Johannes]

Hi,

the colleagues have been faster with the release then me.
2) is on the way
3) There is nothing special about that. If the classification prompt is there, it checks for the content and checks it against the prompt.
If a classification is triggerd, then the changed values will be shown in the widget. If there is nothing to classifiy, it simply wont show anything
at the moment.

So the setup itself (Prompt, content and history) would be of interest.

note: You can look up the classification / routing in the llm audit module.
4) This is on the list to check. Do you have very large FAQs / lot of images?
5) This is available for summary and answer rewrite. As we do not just have "mail" it is not that simple.
Title rewrite and routing are "general" features, as both may modify the queues where they are actually used.
The most of the things run before the other filters. And this is intendend.

You can get in contact using the community@znuny.com (German / English) and we can get in contact for 3 and 4 if you like.
If you have a good suggestion for 5, let us know there.

Regards
Johannes

[root]

Hi,

For #5 I would go with a queue preference. Easy to check, easy to set up.

- root

[finalbeta2]

4) This is on the list to check. Do you have very large FAQs / lot of images?

Yes, this would be the larger FAQ items. When Znuny sends them to Ollama, they seem to be too large causing an HTTP error on the request. Looks like the request should be chuncked. (and that might depend on model in use etc/not sure).