When the bot searches for categories and generates an output, it only displays a partial category slug for second-level categories. For example, Documentation > Using Discourse, which has a valid slug #documentation:using-discourse, is only shown as #using-discourse. Which of course does not render as a link.
The same happens when the bot tries to create a link to a subcategory – it still fails as it produces a wrong slug in the URL.
It would be helpful if the bot had access to the complete slug so it can accurately display subcategories.
Because an AI bot is really stupid. You must explain that in a slug can be a hyphen, and even give an example. If that model is really hard headed you may list every such category slugs.
It is like 3 years old kid. If you explain everything given simple rules to follow, where is no escape routes or doors to start guessing, it works decently good. Most of times. Well unti someone updates that model and it breaks something again.
I tried. I explained. And tried again. It made the main categories right, but not the subcategories. I just think it is not getting the right source data. So even if it wanted, it could not get the slug right. It just does not know what the parent category is. And of course I can suggest and hard-code the category list in the persona definition, but why does it use “Search Categories” function then?
Propably not. It doesn’t know the structure of your forum. And how could it even know? No one has ever told it to it. If that is somehow an important part of your prompt, you have to explain that too.
Because it uses that piece of information to find posts. It doesn’t use it to understand your sitemap
You are not right. The bot is configured to search in categories and it does so. It properly writes down their list. It properly uses a slug and not the name when I tell it to. But it fails with the slug for a subcategory. Most probably because the function that returns the list of categories does not contain it. That is my point.
I got, but I don’t have that many categories either. I have found that everytime when such list type things, meaning there is few with same structure, like category links, increasing over some, AI starts forget its rules and start to use that data it has collected. At that point it starts missing this and hallucinating rest.
Everything else with summaries weren’t that big success… actually those were quite catastrophic, and I abandonded whole idea.
I don’t have that much demand to get links for categories per se, so I don’t know if new OpenAI models are more stupid than old one. More lazier they are.
Then why does it say “Listing Categories” and why does it print an exact list? It can also read topics and search. It understands pretty much very well what is going on. Otherwise I don’t know how else would it be useful inside Doscourse? I am confused. The only thing I was pointing out that the category list function it can call does not return the full path/slug for subcategories. So it doesn’t use it. Nothing else. And I think that the summaries lack exactly the same from what I can see.
I think we may need the AI team to weigh in on this one and break the deadlock.
The idea that there’s something ‘meta’ about the subcategories that the AI bot is just not picking up on is very appealing to me (based on the time I sunk in trying to do it via prompt ). Unfortunately I’m at the end of the AI spectrum so I have no idea if this is a reasonable assumption.
However I did spot this post from Don that says the workaround proved successful for him:
end of the AI spectrum so I have no idea if this is a reasonable assumption. 
