FlexGPT - ChatGPT with memory, web search, unlimited GPT-4, no subscriptions

Just released a devlog and roadmap with some of your suggestions (link is in the app).

v1.1 - 2nd May 2023

  • :gear: added advanced settings to modify GPT-3.5/4 variables
  • :open_book: improved memory efficiency by discarding irrelevant memory chunks
  • :rocket: reworked memory upload process with batching to improve reliability
  • :money_mouth_face: added new ways to earn credit
    • referred users will give a credit to your account worth 10% of any transaction they make on FlexGPT
    • this is applicable for the lifetime of the user
    • get instant top ups with Twitter and Product Hunt
  • :grin: added a public change log
  • :beetle: little bug and styling fixes here and there

And soon:

  • :brain: advanced memory management
    • filter, view, edit, and categorise all uploaded data
    • search through
  • :credit_card: invoices in account section
    • downloadable invoices/receipts in account section
    • …so you can business expense FlexGPT :laughing:
  • :link: embedded URL reading
    • FlexGPT will answer questions about your URL without having to add it to memory first
    • this feature will increase chat latency if you have it enabled (reading the webpage takes 5-20 seconds depending on its content)
    • you’ll be able to ask something like ‘explain Stripe’s currency conversion policy ( https://stripe.com/docs/currencies/conversions)’ and it will give answers based on the webpage contents
  • :newspaper: news search
    • this will work out of the box as part of the web search functionality, so no new options needed.
    • currently web search only includes webpages so asking about news items will rarely be successful
  • :globe_with_meridians: share chat history
    • publish and share a link to a FlexGPT chat
  • :alien: code block formatting glitching
    • when code blocks are being typed, it kind of spasms
    • the fix is a bit clunky but should be improved in future
  • :frowning: GPT-3.5 pricing change
    • FlexGPT grew a bit faster than expected (sorry about the launch day server issues…)
    • GPT-3.5 will cost the same as OpenAI’s API pricing - just $0.002 / 1K tokens. It’s basically free, so this only really affects non-paying users who only receive $0.05 credit.
4 Likes

Getting an error whilst chatting… may be temporary but thought you should know

A couple of ux, ui things.

In mobile it would be nice if the sidebar hid when clicking a button action.

The links in the chat could be better served in a new tab, window.

@stuart4 - This should hopefully be fixed now :crossed_fingers: . That error is from a plugin that I built, so the fault is mine there :grimacing: . We’re working on scaling the backend now due to the huge amount of traffic FlexGPT is seeing. Sorry @georgecollier and crew!

1 Like

Adding the entire bubble.io manual to the memory on mine and using it as a personal app builder guide. Loving it so far! However, is access to memory on mobile not a thing yet?

2 Likes

It’s the same on mobile as it is on desktop (although, the mobile site is a bit bad right now and could be improved…)

1 Like

Are you looking for any help from other bubblers interested in volunteering time towards this project? After how powerful I’ve seen flex to be, I’d be happy to give away time to help out.

@georgecollier, crazy idea for FlexGPT: Add the ability for users to create and program their own bot, just like ORA.sh. Go check it out. I made a bot that’s prompt makes it take on the persona of Dr. Phil, use real therapy tactics on the user in conversation, and then send the user a 7-step care plan with free online resources. Go check it out: Dr.Phil_Bot. But if you do this, and give users the ability to input their own Openai API key, they can program their bots, call their APIs and use FlexGPT’s Memory function to upload info on the topic the bot will be programmed for. That would be super powerful. The memory function is key that’s where FlexGPT wins.

small bug, when updating the categories in memory, looks like its updating the first row… and not the item… even if it’s actually updating it correctly.

chrome_j487DxRhc4

i would like to understand the “Category” drop down… in terms of the AI understanding about 'its storage" and any what constraints it has around the category selected.

will it understand, should it understand ’ list all documents in storage’ ? should it understand what the documents are, in this case resumes.

thanks for the help :slight_smile: 7 days (ish) in to AI now and learning as I go what the rules are.

1 Like

This is an excellent point that @stuart4 brings up. Also, it is implied by the use of the categories filter, that the bot will understand all of the saved material in that category. Does it do this or just know the most recent one?

Hey Stuart! This is because when you edit anything except the title, we essentially delete the old memory item from your database and make a new one (the new one appears at the top as it is most recent). I’ll look into seeing how I can improve this so it makes sense!

Hey Stuart (and @hullaustinj )

TLDR; it just works. Don’t instruct FlexGPT to use memory or categories. Prompt it just like you’ve included the relevant document in the prompt itself :slight_smile:

I am working on improving the documentation for new users so everyone understands how best to use FlexGPT. You can see the WIP guide here, although if any of you would like to suggest improvements to the guide, by all means do so: Notion – The all-in-one workspace for your notes, tasks, wikis, and databases.

Just to quickly explain categories: when you are in a conversation and memory is enabled, there’s a dropdown that enables you to choose a category. When a category is selected, FlexGPT will only find items from that category. A use case might be if you have resumes for one candidate and resumes for another candidate, you could categorise these into two categories, and then choose a category in the conversation so that FlexGPT doesn’t mix up the two. You can change the filtered category during the conversation (just like the model, memory settings, and search settings).

In regards to FlexGPT’s understanding of its memory:

  • FlexGPT does not know what memory it has saved
  • FlexGPT does not know the titles of memory
  • FlexGPT will not understand commands to access certain documents/memory items

Essentially all that happens when you send a message, is we look for sections of your memory (obeying any category filter if you have one enabled) that are similar in meaning and would be useful to FlexGPT when responding. We then give GPT-3.5/4 these sections of memory as additional information so that it can use it in its response.

So, if you say ‘list all documents from category Resumes’:

  1. We put into numbers (vectors) what that statement means
  2. We search through your memory (obeying any category filters) for text that means something similar
  3. We don’t find anything because nothing in your memory means anything close to ‘list all documents from category Resumes’

Instead, you should say something like ‘how can I improve my digital marketing resume’:

  1. We put that into numbers
  2. We find sections of your resumes with similar meanings which in this case will mean sections of text that are more able to be improved, (as decided by OpenAI’s embeddings algorithm :sweat_smile:)
  3. We return these sections to FlexGPT
  4. FlexGPT offers advice to improve these sections

Hope this helps. Feel free to ask away for any more clarifying questions.

1 Like

@georgecollier , FlexGPT is awesome! Many thanks for all your hard work. I am wondering if I can suggest you add folders to categories, as my memory page is filling up, and I know folders and subfolders may help with that. Or even just filtering memory results by category and adding pagination. Just a suggestion.

would love to be able to pay with bitcoin and/or ethereum.

Great stuff! After I have trained the model on the material I uploaded, can I connect the API to my own chat site? For example, let’s say I want to release a chatbot trained on how to start beekeeping in the UK. Can I train in FlexGPT and deploy elsewhere?

Not currently but I believe @georgecollier and team have that on their radar. Follow the Development Log.

Hey @georgecollier , the new chat button seems to be down on mobile. Can you check the workflow for that button on the mobile page, please?

Hey Austin, doesn’t seem to be an issue for me on an iPhone (or on laptop with browser window sized to mobile). Have you tried clearing caches? Also, you can’t open a new chat while a message is loading.

That worked, thank you! Is there a wat you guys could add a stop generating button for when gpt starts going in the wrong direction? That way we can more quickly get it back on task.

Hey @georgecollier , I wanted to share this article with you. While it is not specifically referencing flexgpt, it is referencing what you are doing with it in giving more capabilities to the AI before the big companies even do. That being the case, I wanted to congratulate you for your work and let you know that I personally have canceled my openai subscription and am all in on flexgpt and any open source ai such as huggingface. @georgecollier, u and ur team are leading the way to the future and we support you.

Here is the article: Google AI Documents Leak "We Have No Moat, And Neither Does OpenAI" - Natural 20

2 Likes

Thanks Austin! But there’s no team, it’s just me for now!