AI personas aren’t changing as expected when PM

I have several personas in use. Those are using same model, GTP-4o, and all of them has setting using default enabled. So I don’t see that very confusing menu of different LLMs. All personas are used via PM by robo-icon.

Sometimes, quite randomly, I get almost normal PM window (choose user, topic, tag), but I see last used persona too. If I don’t give user to send, and I trust on persona I see, sending that PM fails. Persona disappear and because receiver is missing. Closing, new try and everything is back normal.

But what happens every time is when I try to change used persona. Last used is always visible. I change it and then popups this (sorry language)

It’s warning of user that has’t seen long time. But user is the very first persona I had. It is never same where I changed, but always that Deckard guy.

I close that warning, write my question and send it. I get answer, but answering persona is not what I was selected, but the last used — and in most of cases other than that Deckard. Then I make a new try, and it works.

So I have to change used persona every time twice to get wanted one.

And one more thing related to that.

If I hide/drop down that first PM-try and open it again, used persona is missing.

This all started today after upgraiding. The version is 3.4.0.beta3-dev (7f607699b8)

iPad with latest OS, and DiscourseHub. Now, I haven’t tried Safari, because really often the composer shows only one or two rows. But I’ll make a topic about that bit later.

2 Likes

Any chance you can make a video of the issue , it will make it easier to debug

1 Like

I’ll try to catch it up that way the issue for you guys would be more obvious. But that you are asking video tells that problems are just mine, not some global thing. My setup is quite typical, though.

But that much I know for sure. It comes from setting Always use default Language Model. When it is enabled it looks like changing a persona will expect very first created persona because of its pseudo-user. If that is disabled everything works just fine — but then everyone must remeber change LLM if used different one; in my case it mainly means alternating between two models, GPT-4o and o1.

I had bad feeling everything comes from very beginning how I handled pseudo-users when the first time LLM was changed. Because at that time visible user changed too at that was a pain UX wise. First you learn to use Bart and then it changed to NewBart, if you understand what I mean. Perhaps I handled it totally wrong back then and that is hounting me now.

But because one setting makes things harder I’m blaming it before I start a bit of a hard job by cleaning everything away and starting all from scratch again.

1 Like

I have a repro of this bug, will get it sorted.

1 Like

This will fix it once merged, also cleans up and clarifies the UI a bit

2 Likes

Thanks! That did the job.

1 Like