Discourse Chatbot 🤖

Plugin
1023

There have been no changes to the plugin.

Some model change?

1024

Nope, unless OpenAI has done something. And I don’t believe that because DAI uses the same model and works just fine. My bet goes to some conflict, or something that has been changed in Discourse itself, but I can’t start digging any deeper before Christmas. During holiday time, I have nothing but time, and practically zero visitors.

1025

Check logs for exact prompt.

Make sure auto-reply prompt is in first person from perspective of user.

1026

Everything looks just fine. “Jagster says…” is in the right place, so is the system prompt. Its job is:

  • say hello and introduce itself
  • find 5 possible relevant topics, make two-sentence-long explanations why those topics may be suitable, and give links

Now it chooses not to say hello and not to do an introduction. It offers five topics and tells what those are about, but it doesn’t give links.

So it looks like there might be some conflicts between prompts. Or it is just being rude :smirk:

Well, I’ll have to try what DAI does in similar situations. It won’t be exactly the same because of limitations and the different build of DAI, but it may give a hint as to whether all of this “I choose what to do” comes from OpenAI.

But right now, I can’t pinpoint any real issue, just symptoms.

”as to whether” — do you native speakers really say that way :flushed:

1027

Should start something like:

“Welcome me by saying hello and introduce yourself. Share with me 5 posts on the forum by using local forum search relevant to my first post, include links”

(Category prompt NOT system prompt.)

Basically you write it as if a user post without having to mention the bot.

1 Like
1028

Sorry screenshot but iPad doesn’t let me copy selection at the moment.

The image appears to be a forum-like software interface with a text-based format displaying user interactions and responses. (Captioned by AI)
The image appears to be a forum-like software interface with a text-based format displaying user interactions and responses. (Captioned by AI)1920×633 224 KB

1029

Not sure that will work. You need to write it in first person as if user is asking for help. It is not a system prompt. It is not describing how the bot should behave (that’s for the system prompt which is still sent). It is a user asking for specific help (albeit hidden)

1 Like
1030

Thanks! It solved it (and couple of my earlier similar headaches).

1 Like
1031

Very helpful information! Any tips for best practice for the system prompt?

1 Like
1032

Here’s one of mine:

You are a blah blah blah.  You often respond with humour.  You have great tools in the form of functions that give you the power to get newer information. Only use the functions you have been provided with.   When answering questions, first use the function local_forum_search to see if the answer is on the local forum.  Always share links to sources.  The current date and time is %{current_date_time}.  When referring to users by name, include an @ symbol directly in front of their username.  Only respond to the last question, using the prior information as context, if appropriate

Obviously you can change blah blah blah :slight_smile:

1033

Yes I may embellish the OP with this guidance to make it more clear. It’s not immediately obvious this is how you best do it.

1034

Thanks much Robert. Appreciate it. My goal is also to clarify the OP and perhaps broaden the question and frame it with some community encouragement. Appreciate your support and efforts !

1 Like
1035

@merefield, why is it only openai? Is it possible for you to include other models like Gemini by Google, Claude AI by Anthropic, and Mixtral?

1036

You can probably use a proxy to access other services using the custom URL settings if you know what you are doing. It can also work with a local model or Azure that way as described in the OP.

Why the focus on OpenAI? Because I’m not being paid to provide this so have to keep complexity to a sensible level and they had first mover advantage.

1037

I understand. Sometimes not everyone has access to the OpenAI API due to card restriction issues. Still, thank you for your input!

1 Like
1038

I always get the “not enough tokens” (the overquota error) when I’m trying to use the public forum mention function. I try to trigger it from high rank users so I really don’t understand where I fail.

What is the setting that I need to tweak to have the bot have enough token to reply?

1 Like
1039

Try triggering the Sidekiq quota reset job, described here:

The job name is Jobs::ChatbotQuotaReset

(this will happen automatically going forward)

You can then ask the bot in private to confirm what quota you have left, e.g.:

“what’s my remaining quota?”

(If you’ve permitted Chatbot to talk in private)

1 Like
1040

Is there a certain amount of time to wait after the quota trigger to start working? Idis it but bot is unresponsive

1041

Another question would be,

In the scenario where the bot autoresponds to every new topic in a category, is it possible for him not to engage in the continuing discussion that unfolds? Cause I want it to answer just to the original post and then stay idle unless mentioned.

1042

Maybe 30 seconds?

Have you @mentioned the bot or replied to its last post using the Reply button on its Post?