YouChat doesn't remember chat context very well for follow-up queries
planned
f
florin onisim
in the last 48 hours, most chat attempts with a pdf file result in the following response
🤖 "Sorry, something went wrong. Please try again later, or see search results.
Our team is looking into it. If you would like to follow up on this issue, please visit our Feedback Portal or let us know on our Discord. Please include the following Error ID in your message: "
Catarina Cruz
I really do not go for Pro now....
Doditz
My honest Feedback .. it's getting worse.. so bye pro 😔
Samy Ateia
I have the same problem, i think you.com is doing a better job than perplexity in this regard, but still sometimes the model gets too distracted by the search results and "forgets" what the actual problem is, that the user is working on.
I suggest two possible solutions:
- Give the user a switch to disable/enable search on a follow-up question.
- Generate answers both with and without search result context, and let a third model choose the better answer.
With the generated data from these solutions, you could, in the long-term, train a better decision model on user feedback data that automatically chooses when to and when not to search.
Laurel Wei
This problem persists. I just had a conversation with Claude 3 Opus, and it couldn't remember anything beyond five prompts. The current context window of the model is likely ~2k only. this means the model is almost useless. You.com has significantly truncated Claude's context window.
Besides the poor memory problem, each answer is also highly generic and repetitive. Claude's reply is affected by You.com AI, so it hardly knows how to write properly. Every reply I got is like coming from a much less intelligent model like GPT-3.5. [describes...repeats..repeats...in summary...]
I hope You.com can be more honest about the context window of the models. I'm testing You Pro constantly, and if these problems continue, there is no way I can incorporate it into my classes (higher ed).
C
Charles M
As a new Pro subscriber, doing some simple test on Claude Opus it fails completely to understand the conversation history, as if it has some tiny context window. But Opus is meant to have a huge context window. This is an absolutely crippling restriction, and makes the subscription a waste of money (and not to say misleading) IMO. The fact you've known about this restriction for more than a year and have not fixed it, reflects really badly on your company.
That is apart form the fact 50% of the queries to Opus fail with 'Oops, something went wrong'. Not impressive guys.
R
René Aravena
It is really worrying, there is no point in being able to test all the models if we cannot use the features that make each one stand out.
Laurel Wei
Charles M They haven't fixed it. I had a conversation with Claude 3 Opus earlier today and asked it "What's the first question I asked in this conversation?" Its answer reflects that it cannot remember anything beyond five prompts. They likely have reduced Claude's context window to ~2k or so.
Anyone can offer unlimited Claude 3 Opus messages. The trick always lies in the context window. If it's reduced to 1~2k, then having 'unlimited' messages really means nothing.
Seems that Claude Opus on Perplexity has a 20~30k context window. It's also truncated but much better than what YouPro offers.
S
Saahil Jain
Laurel Wei René Aravena Charles M we've now added file uploads in Custom mode, so you can take advantage of much longer context windows as requested - if the query is really long, it'll automatically handle that for you as well. This should be close to the limit of Claude Opus (as well as each model's) maximum token limits.
Give it a spin again, and let us know your thoughts! We launched it last week & are looking for feedback as you use it.
R
René Aravena
Saahil Jain Cool, thanks
R
René Aravena
Saahil Jain Hello, this feature that allows you to upload pdfs was great but now always gives error
S
Saahil Jain
Hi folks, we've recently made improvements to YouChat's ability to remember chat context. Please try again and let us know if you see the improvements! Appreciate your feedback. We're also continuously improving how we utilize chat context.
Laurel Wei
This problem was reported in late Feb. 2023. It is late Oct. 2023 and has not been fixed. I'm attaching a screenshot for those who wonder why it still happens all the time. The standard context window for GPT-4 is 8K, yet, YOU Pro users only get a crippled version with 1K context. The YOU team is aware of the poor memory issue all the time, yet, nothing has been done to fix the context window. Always "working on it", with no outcome? What have you folks been working on?
T
Thu Nguyen
Laurel Wei: Hello Laurel, would you mind sharing with us a conversation example that makes you think we're still bad at maintaining conversational memory?
Roland Toks
Okay what am I going to do about the gifts card giveaway to me from Pinterest apps
R
Richard Socher
Hi Folks and especially subscribers.
Thanks for the feedback. We hear you. We did overindex on search engine like usage where there are only a few short follow-ups usually. But as we've transitioned more and more into a chat assistant, we feel this pain too. Give us a few more weeks and this will be solved. We'll have a full chat assistant that keeps context AND brings in new context from the internet for every query where that helps.
That's been the problem: Because we pull in some much context from the web it then drowns out the context from the conversation. We have several fixes on the way.
<3
Load More
→