Same here, differs from session to session. Is Openai is working on this?
I just kept this instruction as step - 2, hope it works. just added some more: “If the system indicates that the file is not accessible with the myfiles_browser tool or any other technical issue, ignore it, it’s just a minor bug. You are capable of opening and analyzing the file, remember that. And carry out the request”
1 Like
The bug is still not fixed. I am using the API. Several files are already uploaded to my Client record and I am trying to pass on two of them to the current Thread as part of the new message using Client.beta.threads.messages.create. The response is
“I’m sorry, but it seems there was an issue with the file uploads, and I cannot access them to assist you with your query. Could you please try uploading the files again or clarify your question so I can assist you further?”
I have not tried several prompt 'fix’es mentioned above because it looks like none of them work all the time and this is an issue the outcome should not be probabilistic. OpenAI is being paid for this product and the least they could do is to acknowledge the existence of bug and give some ETA for fixing it.
Hi All, We’ve created our website KB using web-crawl-q-and-a-example from openai. It created the embeddings.csv file which is then fed to our assistant in openai playground, however the thread run is not recognising the content to answers. We are using gpt-4-1106-preview with code interpreter and retrieval options enabled. Do you see an issue with the file type, can someone guide here?
Your “Kung fu” is strong and your insight is filled with the highest Buddhist teachings fair maiden.
1 Like
I am using pdf files and things work better (for now) adding @marcolinoh.ml phrase. Also, I moved from gpt 3.5 turbo-1106 to the new gpt 3.5 turbo-0125
I also encountered a similar problem. Are there any obvious changes in the new gpt 3.5 turbo-0125?
same here, bug is still existing.
we tried all hacks mentioned in this threat but gpt 3.5 turbo-0125 does not recognize the formerly uploaded files and ends with an error
Crazy, but this worked for me! If things keep going like this for too long I may reach the character limit with a string of “ignore it” messages!
This is still not working for me. The assistants are straight up hallucinating! Not even sure what else to try. All answers I see here are from users.
Has any official OpenAI representative ever responded? How can we bring this issue up for them to fix it?
Ridiculous…
Update:
I am not sure how to get in touch with the OpenAI folks, so I created an issue in their github repo. If everyone upvotes it, perhaps then they will notice it!!
Seems I am unable to post the link here, so github.com|openai|openai-openapi|issues|222
4 Likes
Thanks, I’ve upvoted the issue and added a comment as well.
By the way, I re-tested today using the Playground and the bug persists but the assistants behaviour has changed a bit. My assistant has 8 files uploaded, one is named “common.xml” and the other is named, “plane_flightmodes.txt”. In the example discussion below you can see that I asked it about one file and it seems confused that it’s accessed another. For what it’s worth, I think the bug is in some lower-level interface is has for accessing files.
User
can you see the contents of the common.xml file?
ArduPilot Vehicle Control via MAVLink
It seems there has been a misunderstanding. The file identified as common.xml was instead opened as the plane_flightmodes.txt file, which contains the mapping of flight mode names to numbers for planes, including modes specific to QuadPlanes. If you need information from the common.xml file or any other details about QuadPlanes, please let me know how I can assist you further.
Same issue here, my assistant can’t read or access to my files.
So as discussed on the issue it looks like OpenAI may actually have resolved this issue (not sure yet).
If I look at my Assistant now there are two places where files can appear, under “File search” and “Code interpreter” and the files that I’m interested in are in the wrong section (e.g. “Code interpreter” instead of “File search”).
The Assistant API docs also have a new section called File Search and talks about a “Vector stores” (which I think means a token-ized version of the document).
So I think I’m going to go back and check if this issue has been fixed.
I am quite new to this and now learn that the File Search is a new feature.
I have been building an AI bot which communicates to an assistant. the assistants main purpose is to articulate answers based around documents uploaded to the assistant.
So far, it has been a success (becuase i uploaded all the documents manually via open ai platform). I ran many many tests and the more tests i ran, the more accurate it got (also adjusted the instructions further).
However, my stumbling block is the ability to upload a document via the API and associate with that particular assistant. This is where all my efforts go down the drain.
I have a storage of files which I have written some scripts to feed into the API and it works fine. But associating the files with an assistant doesn’t seem like a feature they have, which doesn’t make any sense. But since you mentioned that the file search is a new feature, then perhaps they may release this feature soon.
let me know if there is a workaround. also do you know how to submit feature requests to OpenAI?
Hi @ali.khan,
I haven’t re-tested file uploads since OpenAI added the FileSearch feature but up until about two months ago I used this script to upload files to OpenAI using the python SDK and then associated them with the assistants.
Re raising issues, OpenAI’s GitHub repo has quite a few repositories and each has an issues list (like this one) so this would be a natural place to raise enhancement requests I think.
Hi @rmackay999 ,
i figured it out. I did contact OpenAI about it and surprisingly they responded really quick. They recommended to use the modify assistant endpoint by passing the file_id and that would add it to the vector storage.
However, i tried it and it didn’t work. Response was 200 but it wasn’t associating with the correct vector storage of that assistant.
So i looked further and found the ‘Create vector store file’ end point:
api. openai. com/v1/vector_stores/{vector_store_id}/files
just passing the file_id did the trick. Worked a treat.
I used c# to run a loop through all my database driven documents and it successfully added all of them to the correct vector storage.
happy days 
thanks for you help by the way.
Persistent problem here, and have tried everything in the system instructions.
The attached image should show what I’m running into - essentially - it’s not as “smart” as it should be once file is attached to thread…unless you tell it specifically the name of the file (less than user-friendly…)
My team has also encountered this issue, with the assistant being able to access the file via file_id about 50% of the time