😬 Fibery User Guide chatbot

As the first step to custom chatbots based on your data, we experimented with a chatbot that knows Fibery User Guide. You can play with it here

https://the.fibery.io/@public/User_Guide/Start-Here-Fibery-User-Guide-6568#ai-assistant/guide

Try questions like

  • How to create a formula?
  • How to create a formula that adds two numbers?
  • What is space?
  • How automation rules work?

It would be interesting to hear some surprisingly good/bad answers here.

We are also thinking about Fibery API chatbot, it seems it will be able to write decent things for you…

5 Likes

Awesome :star_struck: Asked some basic stuff, that works pretty well!

It is a bit slow, some patience is needed.

And I’m still doubting if it’s not secretly @ChrisG PT :face_with_monocle::joy:

1 Like

awesome. are you planning on expanding this to make fibery able to write copy based on the data of a knowledge base?

Can you expand the idea please?

a huge downside of chatgpt is that it’s not possible to feed it a lot of data.

it would be amazing if the AI of fibery would be aware of the data in the databases. (or could be “trained” on data) so that it is possible to use fibery ai to write texts based on the data.

either on data of the current entity, a defined database or on all the data in fibery.

this way we would be able to use fibery to help us write applications for grants, answer questions from users and journalists. bring up knowledge hidden in meeting notes and other documentation.

another usecase: it would be great to train the AI on all outgoing emails to get the “voice” of our communication.

there is probably a lot more stuff that would open up. right now for us it would be a great help to write grant proposals. as we have all the info in out database - but filling out forms with various length requirememts is taking forever. also for new team members to get the info right is usually a big task.

grantable.co is a great example how the ai could be used on a given set of files/entities

3 Likes

This can also be very helpful for PKM use cases.

Plus we have an extra use case. We’ve built the wiki of the workspace in the workspace. Would be awesome if you can have your own AI chatbot for that specific database.

1 Like

Hey there, he’s referring to generative ai. That is, using ai to create text (or images, or other media) whereas currently your use case is as a chat bot (e.g. tell me about xyz).

The generative ai side of things is used a lot in a lot of apps: jasper, surfer, frase, but also some of your direct competitors like coda and notion. And also getting more and more built into Google Workspace and Microsoft Office. Think of it like this - instead of “hey, how do I create a formula to automatically add X +Y,” and instead think “help me write the page of our user guide that deals with X + Y,” and even moreso - particularly able to improve over time - “use my site’s history and reference xxx codebase and write a great article that instruct one how to deal with X + Y” and even more - “automatically write the outline, incorporate 45 LSI phrases, reference six sites from authoritative sources, add the 12 most asked questions and provide the answers, and incorporate EEAT so that - even though content marketing is bad and doesn’t work (hehe I read your blog) - it will work in this case.”

This of course can be expanded into emails, social media posts, Q&A, etc. - lots of companies are doing that - but - considering your model - providing that out-of-the-box is a natural fit and would help a lot of startups - particularly any that may try to use Fibery as a CMS. Also, we are approaching a point (if not there already) where it’s less about a differentiator and more about a perceived shortcoming if you don’t have it - particularly given how aggressive MS Office and Google Workspace are going.

In terms of some of the other comments I see, they make sense and you should look to incorporate them into things early-on. Some of them, plus some other things:

(1) Don’t limit yourself to openai. Maybe at the start, but there are others and the open source LLMs coming out of Meta are particularly interesting. So, too, for programming and development, are some of the models out of hugging face and also particularly Phind.
(2) Keep in mind there is a lot of open source stuff that is amazing right now, particularly stability.
(3) Another post mentioned space-specific fine-tuning. I agree with this. But first I will say that you should pass along user-enterable custom instructions. These need not be complex or complicated but can take out-of-the-box openai from a 5 to an 8+ in about 10 minutes (it’s as simple as two text fields). What’s more, if done right, they need not be a one-time-thing but instead could be per-space, per-database, per-personna, per-need, etc.
(4) From there, you can enable fine-tuning but that’s a bit more effort; however, there is an innovative way to do this that is straightforward - though I’ll have to write a full post on it - and could use some feedback about your tech stack first. Fine-Tuning enables you to “train” in a much more specific way than merely custom instructions. You can train data, differentiation, tone, style, etc. So your customer space could have one particular voice while your marketing space could have another and your product development could have yet another. Once you’ve evolved, you come to appreciate just how powerful (and surprisingly simple) fine-tuning is.

Feel free to reach out regarding this, as I have used Fibery I’ve naturally been considering how and where AI could (and should be integrated).

Hmm, we have this in Fibery already, just want to make sure you did not miss it

https://the.fibery.io/@public/User_Guide/Guide/Fibery-AI-Assistant-in-Rich-Text-and-Documents-156

This post is about chatbots, not generative AI. The feedback is relevant for generative text assistance though. We don’t lock ourselves with OpenAI, but so far it is the best provider in terms of simplicity, speed and quality.

1 Like

I imagine these chatbots have been trained on the text contents of the Guides. Are they also trained on transcripts from the Fibery YouTube videos? That might be a nice addition.

I have always felt that the Fibery docs/guides are a little sparse when it comes to specific methods of accomplishing things, like “is it possible to create one-to-many recursively nested rows in reports (the same as in list or grid view)?” or “has anybody else ever written a formula that already does this?”.

I have an idea rolling around in my head about the chatbot using its responses and feedback the users give it about its responses to generate new documentation and instructions on the fly, which is injected into the existing documentation. Then, the whole system and its contents are refined based on user feedback about the documentation itself, and/or based on how frequently users visit certain documentation pages and seek additional clarification about related topics from the chatbot.

I haven’t ever actually built anything with ML, so I have no idea how feasible or useful that would be. But it seems like a useful concept. It would probably even mean the documentation would be resilient to errors in the chatbot-generated instructions AND resilient to documentation getting out of date as the product changes, because the portions of the outdated documentation that users reference the most would get targeted for revisions by the chatbot as users test the outdated documentation, don’t solve their problem, and turn to the chatbot for additional help for that topic.

Super excited for Fibery-built documentation chatbots though. Amazing feature, cool possibilities.

1 Like