[DONE] GPT3 implementation in Fibery (share your ideas and cases)

Some inspiration for use case

Almost all of this will be possible in Fibery next month (or sooner) when we will add GPT to automations :slight_smile:

2 Likes

How to get and set OpenAI token — Guide | Fibery

:eyes:

1 Like

Some markdown support

2023-02-10 16.01.37

4 Likes

We have some version that you can play with already. It is not ready for the public release, but still it will be interesting to see what use cases you can invent. Here is the guide

https://the.fibery.io/@public/User_Guide/Guide/Fibery-AI-assistant-156

Main things:

  • You will need your own OpenAI token. It is free though.
  • You can create your own commands and execute them on selected text or the whole text field
  • You can use templates in commands and access entity fields like this: {{Name}}, {{Feature.Name}}, {{Description}}, etc
  • You can ask to generate results in markdown and insert the result into a text field as a markdown
  • You can make commands private (visible to you) or public (visible to all users)

If you want to play with Fibery AI Assistant (we should invent a better name), :heart: like this post and I will write to you in private how to access it :slight_smile:

NOTE: there will be no default commands, you will have to create your own prompts and save them as commands so far

P.S. Automations are not there yet, but we are working on them already.

8 Likes

Woo this is amazing :slight_smile: Can’t wait to play with it.

1 Like

Okay, here comes the feedback @mdubakov:
Critique sandwich:
The not so great parts:

  • I noticed that the AI command modal appears on the top of selected text. When text is long it’s confusing as I select the text and need to scroll all the way up to see it. Would be better if appeared next to cursor.
  • Would be curios to see if Fibery can do auto-splitting based on headings?
    • I tried to summarise my notes for webinar and they ended up being too long. Maybe AI assistant can calculate number of words / sentences on fibery and split long text accordingly into multiple requests to openAI.
    • The error I got - (400) This model’s maximum context length is 4097 tokens, however you requested 5928 tokens (3880 in your prompt; 2048 for the completion). Please reduce your prompt; or completion length.
  • For referencing other fields in entity would be great if we have some autocompletion or at least validation
    • It’s very easy to make mistakes in just plain text and with all those special characters {{}} I feel non programmers will struggle a lot.

    • An idea I had would be to provide on hover preview for substituted fields - when user hovers cursor over {{field}} it will show preview for that field. That way it’s relatively easy to verify that it’s the right field and makes it easier for non tech users.

Epic parts:

  • Being able to reference data directly from fibery opens so many possibilities. Especially, when automation support will be finished. I can’t wait. Already done semi-automatic summarization of webinar notes.
  • I love idea of commands that can be shared with teams essentially creating reusable prompt interface specific to domain that the team is operating in.

Questions:

  1. I didn’t understand how to reference things which are not linked to a specific entity. For example, let’s say I have intercom chats integration enabled. Would it be possible to get AI assistant to automatically figure out which features user mentioned from the intercom chat log itself based on set of features that we have in the system? If that so how do I reference “all the features in the system?”
  2. About privacy of commands - when I’m logged in as the same user I cannot see private commands created by the same user. It looks like they are private to specific computer. Is this intended?
5 Likes

I’ve tried it a little bit and to add to tpaktop’s comments I just have one of my own so far.

Enter/Shift+Enter/CTRL+Enter
I made the mistake several times of pressing Shift+Enter expecting a new line and Enter to run the command. But it seems it’s the other way around for now. I believe the standard is SHIFT+Enter for new line and CTRL+Enter or Enter for executing (probably a bit different between different software).

Question
Does everyone who wants to have the AI Command visible need to add ?ai in the url?

Future
Could you perhaps tell us a bit about what your plan is for the future? Is your intention to be dependent on OpenAI (or other services) or are you planning on doing your own implementation using GPT as the language model? I’m probably exposing my ignorance here but now the question is asked :slight_smile:

I’m asking mainly because we (and I expect many other users) may have sensitive data that they do not want to share with OpenAI, especially if prompts are being used in their training set. One use case many will probably have is summarizing meeting notes. And if the meeting notes contains sensitive information I would be less likely to use that.

I would be much more comfortable knowing that you/Fibery had all the information stored here and not shared. Or at least not allow the data to be used (although may be hard to enforce).

I found this for Notion AI and I see they use OpenAI in the background. I’m not expecting a similar document now from Fibery but as I mentioned it would be interesting to know how you see this going forward.

2 Likes

I find the following quite interesting:

Notion -

“Notion uses technology provided by OpenAI, LLC (“OpenAI”) to provide Notion AI.”

“We do not and will not permit third parties to use your Content or Customer Data to improve or train their AI models.”

OpenAI -

“As part of this continuous improvement, when you use OpenAI models via our API, we may use the data you provide us to improve our models.”

It’s not clear to me if that means customer data in Notion is (or may be) used to improve the models or not… :thinking:

4 Likes

That privacy question is a really good point. Despite my general distrust of GPT-3, I have really warmed up to the idea of using it in Fibery. But if everything I send it were going to get incorporated back into OpenAI’s training corpus I think that would be a dealbreaker. When messing around with GPT-3 and asking it about my (fairly niche) research areas, I’ve had it spit Wikipedia articles back at me verbatim. Which makes me think that if it were trained on some of the research notes I keep in Fibery, it could potentially also spit some of those out verbatim.

3 Likes

Not yet, this is mentioned in the guide. “Fibery AI Assistant is not trained on your Fibery data (yet), it works only with what you send it in a prompt.” We do want to make it Fibery-aware as the next step.

Interesting, are you sure that this is the case? Commands visibility might be tricky, since some commands are bind to Database.
Can you check that it is indeed devise specific.

2 Likes

I have a command - summarize which looks like this:

{{input}}
Summarize

It’s not accessible in documents within the same workspace. Neither it’s accessible in any other entities. Looks like it bound itself to that specific entity type and I don’t follow why.
It’s just used with generic text input I didn’t reference {{Description}} fields in it as you can see.

Thanks! We’ll see what we can do here.

I think this is the next top priority problem for us, indeed docs can be large and it is a serious limitation for current AI Assistant.

We have this in mind and will try to improve in future versions.

1 Like

Looks like a bug on our end, we will check!

Thanks! Indeed we will fix it.

So far yes. This is a hidden feature and we will make it experimental soon.

1 Like

We are experimenting with this so far. In general it seems we will go for our own GPT implementation to make AI Assistant context-aware and consume all Fibery data daily. But it is too early to say, we are in the very early phase here and want to find some useful patterns first to move forward.

4 Likes

Thank you for the clarification. It does make a lot of sense to start with something already existing. And I’m happy to hear you are considering your own implementation (even though I understand it may not happen in the end).

First steps in Automations + GPT3, extract info and put it into fields

And here I generate some sample data based on GPT3 responses

Batch Actions + GPT3

2023-02-15 20.29.43

Do you have real scenarios you want to run? What are they?

3 Likes

I have some AI automations I’d like to run on the Readwise pages I’m pulling into Fibery from Obsidian.

So if I have the page below, I want to extract each of the inline fields (key:: value) into table fields. I want to link fields like Author to an entity in an Author table.

I also will want to generate an AI summary of my highlights into a Summary field.

Can you copy and paste a text from some page like this here?

1 Like

I’ll probably also use it to generate random items from a list to enter into a field, since that’s missing from Fibery’s native formulas.

I’ll also probably try having a central “tag” database and have a field in other databases where AI suggests tags for entities (and hopefully can connect them to the tag db).