I am having this issue every time. Here’s my response ID:
resp_680eb4e1f2d481929f544b547cd515b406feb2e37e380e46
If OpenAI support isn’t having this issue and @tort_cabbie is also having this issue, could this be an OpenAI permissions issue?
I knew I wasn’t delulu 
With chat completions, the issue is that only the last PDF in a multi-part user message content is employed.
I had the same success on Responses as you, so one has to look further: can the PDF file mechanism supply 100+ pages for vision to the AI - or does it silently error-out?
Since you have your own storage being employed, harder to replicate, I would review:
- Are you using the same project API key for all transactions;
- While docs say “any purpose”, I would use “
user_data” when uploading. - Any other potential complications, such as metadata or parallel vector store use?
━━━━━━━━━━━━━━━━━━━━━━━━━━━◣
OpenAI File Storage Utility ▮▮▶ Select New File 'Purpose'
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬◤
[1] Assistants - docs for search or code interpreter (input)
[2] Assistants_output - files produced by assistant or code (output)
[3] Fine-tune - JSONL training file (input)
[4] Fine-tune-results - learning metrics report (output)
[5] Batch - JSONL list of API calls to run, results (input,output)
[6] Vision - Images for Assistants message attachment (input)
[7] Evals - Model quality assessment file (input)
[8] User_data - New unknown assistants purpose (input)
[9] Exit
◄ current purpose ► assistants
1 Like
Def not delulu. I have the results saved from the same image only pdf summarized beautifully on Monday, only to be unreadable with the same file, code and prompt the following day. Been following this thread ever since hoping they get it figured out. Had to switch to Gemini in the mean time.
1 Like
I did specify “user_data” and am not using any additional metadata or a vector store. Can you try testing my scanned PDF example via base64 PDF input? I can’t find any luck with file upload or base64.
Same! Was working perfectly until last week. My users have been putting my app on fraudwatch due to this issue. I don’t want to use pdf2image and install poppler 
.
yeah, I’m not going to do the pdf2img thing either. Luckily I was using the vercel ai sdk for my project so it was as easy as changing the “model” property to use gemini temporarily. I prefer gpt-4.1-mini, but at least I was able to keep working while I am waiting for this to get fixed.
I shared my code above, super bare bones. I tried @Jackie_Ni’s scanned.pdf file both as base64 and file upload and got the same result with both approaches.
The Prompts Playground sends to the API as BASE64. They also must be data URLs. You can inspect the browser’s network request and parameter use (and steal the body as example and encoding). Look to my prior screenshots to see correct usage (docs are crummy).
Let’s create something you can replicate.
That follows a pattern of “developer knowledge”. Just user messages are capable of full PDF (text+vision) with input_file type, which limits some uses.
Now the user input in the right pane:
You should be able to answer:
- What's Cuban's website?
- What real cost savings does he cite.
Answer from metadata you also receive:
- PDF file names
- PDF page count
- PDF-to-image image count
Then running combined - essentially identical results; questions answered:
Also - update your library SDKs!
1 Like
Hey @_j, thanks for the advice. I tried using the Prompts Playground but its still cooked.
Is it at all possible this issue doesn’t affect all users like yourself but only some?
1 Like
It is certainly possible.
Try store-> false. You might be blocked from jamming up distillation storage with PDFs.
I would look into what self-help is available:
-
Review the default project. That one should be reserved for your “no limits, no blocks” usage. Nothing set in project limits for models.
-
Ensure playground is “your org” → “your default project” at upper-left.
-
For file ID use with API calls, see that dashboard → storage has the PDF with that org/proj, and give a re-upload anyway.
-
Re-issue an API key to the default project. Nothing set for endpoint restrictions. Employ that key.
There really isn’t much else to configure. The playground eliminates the API parameters that were previously documented wrong.
You can then discover if it can only be related to incorrect organization model provisioning, which would have to apply to multiple models for them all to fail, or a stealth pattern of restricted features.
2 Likes
@Jackie_Ni, I was able to get mine working by creating a new project in my account. Dropped your file and mine into it with a simple prompt of “please summarize this document.” and it worked fine. Something must have happened that borked my Default project with one of their releases.
Thanks @_j , didn’t think to look at the project as a possible variable until you started calling out the default one. Figured that was the last variable outside of some hidden user restraint.
OpenAI_Support, hopefully this helps you hone in on the actual issue.
3 Likes
Wait you’re right that worked! @OpenAI_Support can you please investigate why Default project isn’t working ASAP?
1 Like
@OpenAI_Support is there an ETA for fixing the Default project bug? If not, I’ll swap over to a new project for now assuming it doesn’t mess with any rate limits.
1 Like
same, I also meet the issue with pdf file
i’m using azure openai
my request id:
3ff6ff41-02ca-42c7-bc09-355de23f1ad7
8afa8a29-f127-46db-8ef2-e255e8bc0e96
I tried with both gpt-4o and gpt-4.1, chat api and response api, and get the same result
'The server had an error processing your request. Sorry about that! You can retry your request, or contact us through our help center at oai-assistants@microsoft.com if you keep seeing this error. (Please include the request ID 3ff6ff41-02ca-42c7-bc09-355de23f1ad7 in your email.
and when I try to upload file with purpose user_data, it will return Invalid value for purpose.
status: 500,
headers: {
'apim-request-id': '6bec2ff3-bfae-479e-965a-e3c51005b66a',
'content-length': '382',
'content-type': 'application/json',
date: 'Wed, 30 Apr 2025 06:26:32 GMT',
'openai-processing-ms': '562',
'strict-transport-security': 'max-age=31536000; includeSubDomains; preload',
'x-content-type-options': 'nosniff',
'x-ms-middleware-request-id': '6bec2ff3-bfae-479e-965a-e3c51005b66a',
'x-ratelimit-limit-requests': '30000',
'x-ratelimit-remaining-requests': '29997',
'x-request-id': '4b630db8-1139-481d-98a6-a07d7613abf8'
},
request_id: '4b630db8-1139-481d-98a6-a07d7613abf8',
error: {
message: 'The server had an error processing your request. Sorry about that! You can retry your request, or contact us through our help center at oai-assistants@microsoft.com if you keep seeing this error. (Please include the request ID 4b630db8-1139-481d-98a6-a07d7613abf8 in your email.)',
type: 'server_error',
param: null,
code: null
},
code: null,
param: null,
type: 'server_error'
}
Bad news @Jackie_Ni, my new project started exhibiting the same behavior. It was working fine but now it can also no longer read image only PDFs. I had to create yet another project to get it working again.
@OpenAI_Support This is the response ID from the new project with the failing request. I was able to create a new project and drop the same document in and it worked fine.
resp_6813b89e581c81919089c48d9c228187069c9c104773b799
This was working fine a week or ago.
Now it is successfully returning text summaries without image summaries or vice versa, but not both. If you ask for both, it returns ill-formatted HTML and bugs out. We are using the chat completion endpoint and base64 image.
We are taking this out of production as it is unstable. Maybe Sam needs more NVIDIA Blackwell chips…
Just curious are you using structured output? I am doing similar stuff and noticed that it works FINE without no structured output (just give me the transcript etc) but it will fail when demanding structured output. (So I ended up dropping the structured output and just going with pure markdown.
Not using structure output - using HTML.
Interesting - Just tested again with Prompt: “Provide a summary of the text and images of the PDF” - the result was stellar.
If I change the prompt to: “Provide a detailed summary of the text and images of the PDF” - No dice. Got corrupt HTML. Is there such a thing as a detailed summary?
Why not let it output markdown and then make it html afterwards? That is way easier for the LLM