Hey everyone!
Today I launched FlexGPT (FlexGPT - Product Information, Latest Updates, and Reviews 2024 | Product Hunt). It’s just like ChatGPT, but with everything it’s missing:
- Long-term memory - upload documents, text, or webpages
- Internet access - get access to the latest, most up-to-date information
- Unlimited GPT-4 - no GPT-4 rate limits. Enjoy.
- No subscriptions. Pay only for what you use. And GPT-3.5 is free.
In terms of tech, it uses:
- OpenAI’s GPT-3.5/GPT4 for chat
- GPT-2 embeddings for vector processing
- External API for chunking text
- Bing Web Search API
Feel free to give any feedback or ask any questions!
4 Likes
I tested it, good job.
I use chat gpt to code basically, and I love the black box with code inside, and here we don’t have. It would be nice if you can put this in your app
2 Likes
Yes, it would! Trouble is, dynamically inserting Markdown in Bubble is pretty tricky. For now, code is defined by BBCode, but having the embedded code copy feature is definitely something I want to implement in future.
looks like there could be an issue @georgecollier - added credit and its timing out.
Stuart, sorry you’re having this issue. We use a plugin which uses an external server and it seems the server has failed. I’ve contacted the plugin owner to fix this and am looking at running my own server. Let me know if you’d like a credit refund in the meantime, and thanks for your patience 
1 Like
no all good, ill wait… I just like what you’ve done and what youve got together in this app… pretty interesting
@stuart4 @georgecollier - very, very sorry about this. The issue should is resolved now.
1 Like
Dont be, things happen… its all good. Im looking forward to exploring
1 Like
Love what you’ve done with this! Can you add a way to remove the restriction of characters in the input? I would be great it GPT could take an input of much larger sizes.
1 Like
It would also be cool if you added an autogpt model like agentgpt, but instead of just listing tasks, they could be editable tasks. Agentgpt doesn’t allow you to edit the tasks.
1 Like
Yes, I will be adding options to improve token efficiency (for messages older than X chat turns, only include a summary in the prompt rather than the entire text).
1 Like
Trying to use the memory component.
Ive saved a text element that states my name and where im from. However, it doesnt seem to recognise me… maybe ive misunderstood something?
What i provided: im Stuart read, i lm from the uk but live in Sydney Australia
What i asked:
Where am i from
As an AI language model, I do not have access to your personal information such as where you are from unless you provide that information to me. However, if you are interested in finding out your ancestral heritage, there are DNA testing services available that can give you a breakdown of your ancestry. One such service is 23andMe, which offers an ancestry kit that provides ancestral and geographic histories. You can explore more about DNA test kits and how they work on KnowYourDNA and learn how to find out where your ancestors come from using the FamilySearch app.
1 Like
Hey Stuart
This is because GPT has no way of knowing that ‘I’ in your message refers to the ‘I’m Stuart’ in your memory. See below where I added the same memory as you:
One thing to note is that if you have memory or search enabled when they’re not necessary, it can end up giving worse answers. If you know you’re chatting about something in your memory, it is better (in terms of latency, cost, and accuracy) to disable search as well otherwise it can get distracted by irrelevant search results when it already has the answer in memory.
Also, memory currently returns the top n text chunk matches no matter how related they are. So if you have one memory item that’s actually relevant, and two more that you uploaded to another conversation, it’ll retrieve all three. Improving this today
I’m going to be adding a message ‘debugger’ where you can see what information was retrieved to create the response
Cool, thanks… im trying to break it, evaluate it from a non tech perspective… this is who we want to pay in the long term…
This good info.
Also in terms of the bing search, does it read the api title snippets etc to determine relevance… or (and i doubt) does it read the linked page and understand the content of that page.
1 Like
Switched to Google search actually. Bing increased their prices from $0.003 per request to $0.015 per request which I’d have to pass onto users.
But yes, it reads the title, URL, and snippet (because generally the snippet does have useful info).
What I initially wanted to do was when a user includes a URL in their message, automatically scrape it so they can get relevant answers without going through the upload to memory step first. Trouble is, scraping the website takes 5-20 seconds depending on the website contents. Then, vectorising it is another 5, so the latency just becomes a problem and it feels slow. I’ll keep looking into ways I might make it work though.
1 Like
Yea interesting, i think people (depending on what they are trying to achive) would be happy with delayed responses to ‘tasks’. Especially if they were given info on what was happening.
‘Are you happy to wait while I gather live data?’
1 Like
On this , thinking out loud, could you maybe make it part of the prompt rather than a turn off turn on feature in the sidebar or both…
Eg,
prompt: based on memory, tell me where Stuart is from
Prompt: search the web, list the latest reviews of cafe x and tell me if it they are positive.
Ignore me im just making up irrelevant features before I know how to properly use it 
