AI generated Thread Summary, Tags, Prefix, SEO Meta, Thread cover, etc.

[Alpha1]

AI / LLM's can be used to make threads and other content more useful and better organized. For better usability, SEO, content discovery and readability.

One of the challenges of using forums is reading through mega-threads with thousands of replies. Especially if high quality replies and useless comments are mixed. For this reasons there have been many requests for a summary feature over the years. Now with the advent of AI the technology to create good summaries for large amounts of text is possible.

Some other challenges for forums are:

1. Correcting useless thread titles. i.e. 'HELP ME NOW!!!' or 'Dumb question'.
2. Optimal SEO meta tags.
3. Useful thread tags.
4. Correct thread prefix.
5. Correct thread type. i.e. a Question posted as a Discussion Type.
6. Making forum threads look better with images.
7. Making unreadable posts readable. When users post broken English, chat speak, mobile t9 autocomplete suggests wrong words, ALL CAPS, no punctuation, no line breaks wall of text, no formatting, then it can make posts completely unreadable to others.
8. Link to related content from other content types for better content discovery. For example if XFRM has a resource that addresses the topic in a thread.

These can be daunting tasks and many admins don't even bother because of the scale of fixing all that.
Some forum platforms like Discourse and Quora have already implemented some LLM solutions. For example:

Discourse AI Topic Summary 🤖

ℹ Summary Uses a remote AI language model to prepare and post a summary of a topic. 🛠 Repository Link GitHub - merefield/discourse-ai-topic-summary: Uses a remote AI language model to prepare and post a summary of a Topic 📖 Install Guide How to install plugins in Discourse ❤ Sponsorship...

meta.discourse.org

Discourse AI

:discourse2: Summary Integration between AI features and Discourse 🌐 Website Discourse AI Features | Discourse - Civilized Discussion 🛠 Repository Link GitHub - discourse/discourse-ai 📖 Install Guide How to install plugins in Discourse Discourse AI Discourse AI is our one-stop...

meta.discourse.org

Once an LLM understands the meaning of the thread, the LLM can create the following with the data:

1. A thread summary. (widget)
2. A better thread snippet.
3. Suggest a better thread title.
4. SEO meta tags: Meta title, Meta Description, Meta Keywords, etc.
5. Thread tags.
6. Select Thread prefix.
7. Select Thread Type.
8. Thread cover, caption and thumb. This would make forum content look much better.
9. Making unreadable posts readable. (correct spelling mistakes)
10. Display related content. (Widget to display related content from all content types)

And the same for all content types. (XFRM, XFMG, Add-ons)
Please consider to add LLM functionality to let AI / LLM suggest the above.

As there are costs for the use of AI / LLM, it would require various limits on when analysis takes place. For example generate x,y,z only after x posts have been made. Or add a budget for AI credits.
As LLM generated data can be erroneous, methods of correct are needed like community editing of summaries or admin view of new tags.

 

[Sal Collaziano]

Would this plugin be able to talk to a local model via the ollama, openai, or lmstudio protocols?
I insist that we don't transit our data to a third party. So it's a dealbreaker for me if it can't do that.


Also you asked for ideas, here's mine.

LLMs are very expensive to run. If i was designing this, i'd only update a summary for a thread after a user settable number of posts in that thread has been made since the last summary regeneration. By default, this should be 1 page ( use xenforo's setting )

Only ~1% of posts are active at any time, so the daily re-generations would be very minimal.

I would also let the user set a maximum number of summary generations per daily batch job. I'd set this to something conservative like 10 threads per night so that out of the box, someone doesn't rack up a surprise bill with their AI provider.


Another way to make money on this is to resell inference. Allow the user to pay a per token fee to you for convenience, or connect to their local provider.

We have deployed our own LLM using Ollama. We will offer options to use your own local model.

We are taking the potential for the feature to be abused and/or used too much into consideration. There will be options and measures in place to prevent that.

 

[Sal Collaziano]

I think the idea of using AI to write thread titles can be extremely useful for SEO and onsite searches. I also think if implemented correctly it will be very well received by most users.

I think by default the thread title box should be locked. The user should just input their thread message and when the user chooses to post, AI automatically creates a title for the thread, giving the user an option to override it. I think most users will accept the AI generated title, and for those that really want to create a personalized one they can easily override it just before it posts.

Something like this would probably work best for new users. In my experience, they need more help than members who've been around a while and have properly "learned the ropes". In my mind, something like:

"Based on your message, I'd like to suggest the following title for your discussion: [suggested_title]"

With two buttons above or next to the submit button: USE THIS TITLE / KEEP MY TITLE

 

[ProCom]

LLMs are very expensive to run. If i was designing this, i'd only update a summary for a thread after a user settable number of posts in that thread has been made since the last summary regeneration. By default, this should be 1 page ( use xenforo's setting )

Only ~1% of posts are active at any time, so the daily re-generations would be very minimal.

Yup, at scale on a really big / busy forum, this could get extremely expensive really fast. It would also be tricky to quantify the ROI on implementing a feature like this.

One thing we considered (assuming the costs to generate the summaries became exorbitant) was to limit the volume of summaries to only cases where people wanted it. Maybe by providing a "generate summary button"... either for all members, or just for our premium members.

 

[Sal Collaziano]

Yup, at scale on a really big / busy forum, this could get extremely expensive really fast. It would also be tricky to quantify the ROI on implementing a feature like this.

One thing we considered (assuming the costs to generate the summaries became exorbitant) was to limit the volume of summaries to only cases where people wanted it. Maybe by providing a "generate summary button"... either for all members, or just for our premium members.

That's how we're going about this:

I yr anniversary

Next month is my 1st anniversary lol. Just wondering if people are happy with there Seltos? My first impressions were very high tech,fuel consumption great. Now still feel the same but still freaked out when it takes over steering. Was sceptical about no rear aircon vents,but rear seats never...

seltosforums.com

...a Generate Summary button.