ChatGPT: GPT retrieval is failing for multiple users and can be reproduced by me with a new GPT, only making available the first injected document.
Knowledge search failed and skipped with its own chat entry remaining persistent, then devolving to code interpreter:
Additionally, the injection of file IDs for the mount point has duplicate IDs for the same file, with example above.
Hope that can be passed along for resolution.
Here is a workaround I posted as a solution to another GPT problem: I have created custom GPT (my GPT) on GPT-4. How so I run in on GPT 3.5 turbo? - #12 by SomebodySysop
To be clear, this is a solution that requires a developer, and only makes the slightest bit of sense if you intend to offer your solution outside of GPTs (ours was created before GPTs, and we had no intention of using GPTs at all until we saw this post: You can access Mistral API via Custom GPT action)
And, if course, it will come with itās own challenges. But Iāve implemented this, and it works.
GPT <------ API Action ------> RAG Knowledge Base
Now you have complete control over how the knowledge base is embedded and searched. And you can still put it in the GPT store!
Agreed. But, the closest competition I see today is Google ā and they wonāt be winning any customer support awards too soon either. However, letās give credit where credit is due: The āLeaderā system they have set up here works remarkably well. I browse the topics almost daily, and not only are most of my questions answered, I see that those of many others are also answered pretty regularly. Tough work. I donāt know how they do it. But, they do deserve a round of applause!
Your GPT action has you paying the bill for on-demand AI-powered services for the benefit of making OpenAI consumer products more useful?
Now you have complete lack of control over billings from your public-facing free set of instructions.
The only part of GPT retrieval that seemed to be working was the injection of truncated documents, with no relevance of what is presented by user input, which might be only the first one uploaded.
The search aspect that is the rest of this only returns an error to AI.
In my use case, the only users who can use my GPT are users who are already subscribed to my existing RAG service. For them, this is just an added benefit to a service they are already paying for.
Thatās why I said it only makes sense if your application is available outside of the GPT ecosystem.
Well, in my case, I have total control over billings since only paid subscribers can access to begin with.
I donāt know about anybody else, but for me, this only enhances my service offerings. And I donāt have to rely on the GPT embedding scheme for my knowledge base.
Sorry you are correct, I meant "myfiles_browser" instead of "browser". Original post has been edited with the correction. I donāt think it has to do with file IDs, I think the myfiles_browser is not even present in the available tools.
If you inspect element on a previous conversation where the custom GPT was able to use the knowledge search function, go to the Network tab and reload the page. Youāll find two json responses corresponding to the custom gpt and the conversation data, which will be named the same as the parts of the url (https://chat.openai.com/g/<custom gpt>/c/<conversation>).
Looking at the conversation log json response, a successful knowledge search will go something like:
"message": {
...
"content": {
"content_type": "code",
"language": "unknown",
"text": "search(\"user requested query\")"
},
"status": "finished_successfully",
...
"recipient": "myfiles_browser"
},
but now if I look at the json response for the custom gpt configuration, there is no "myfiles_browser" in "tools", leading to the Unrecognized recipient: myfiles_browser error that people are encountering.
Very possible that the immediate double response is result of the gpt catching the missing myfiles_browser and responding using text or python solution instead.
But who knows! OpenAI doesnt seem concerned about it so maybe the system is working 100% as intendedā¦
yep, still " Iām unable to directly search documents at the moment. However, based on standard practices and my existing knowledgeā¦" this is the response I am getting for the past 3 days. Surprising that it goes on for this long without response from OpenAI.
The dedicated GPTs are so great, and my workflow has grown to depend on them for managing a lot of information quickly and efficiently. I hope this gets fixed soon
Iām in the same boat. Itās shocking to me that OpenAI isnāt at the very least acknowledging the problem, given the popularity and level of community contributions. Horrible.
Hey folks, sorry for the delay here, I flagged this to the team, we are looking into it!
Having the same issue with one of my custom GPTs that uses guidelines from the knowledgebase to be perfectly informed to give advices.
I noticed this the other day. I had to frame a question to reread from the start of the conversation. Even at that I was not getting the focused response on the topic as when I first started the conversation. Seams like it has problems when you continue the conversation days later.
It looks like the functionality is back. The PDFs in my custom GPTs are able to be read again.
Yeah, the functionality of reading data from knowledge source is back!
Everything is working again for me too. I can ask for very precise information from the document of its knowledge, etc. Hooray!
Hey folks! This should now be resolved. I am going to close this thread so it is easier to track if things like this happen in the future. Thanks for being patient and building with us : )