"API returned unexpected status code: 413" on OpenAI compatbile OCR LLM endpoint

[CyWS]

Hey there,
been using paperless-gpt now for hundreds of documents with an OpenAI compatbile API from Ionos Germany. Since a few days (can't point it to any update or date unfortunately, last time it worked was 2nd Nov 2025), OCR has stopped working with that error: API returned unexpected status code: 413

My understanding is that HTTP 413 points to "Content Too Large", so the generated image might be too large for Ionos API? Their documentation https://docs.ionos.com/cloud/ai/ai-model-hub does not really say anything on that. I am heavily below the rate limits and I chose a model capable of image input: https://docs.ionos.com/cloud/ai/ai-model-hub/models/llms/mistral-small-24b

Alll the LLM activities Title, Tags and Date work totally fine.

Any chance to further debug this?

Please see the log:

time="2025-11-24T11:33:54Z" level=info msg="Using token limit: 5000"
time="2025-11-24T11:33:54Z" level=info msg="OCR provider: llm"
time="2025-11-24T11:33:54Z" level=info msg="Settings file not found at config/settings.json, creating with default values."
Using paperless-gpt as manual tag
Using paperless-gpt-auto as auto tag
    ╔═══════════════════════════════════════╗
    ║             Paperless GPT             ║
    ╚═══════════════════════════════════════╝
Version: v0.23.0
Commit: fecd76a5bf14f8183b53d874b31b0121981a9779
Build Date: 1758104151
Platform: linux/amd64
Go Version: go1.25.0
Started: Mon, 24 Nov 2025 11:33:54 UTC
time="2025-11-24T11:33:54Z" level=info msg="Initializing OCR provider: llm"
time="2025-11-24T11:33:54Z" level=info msg="Using LLM OCR provider" model=mistralai/Mistral-Small-24B-Instruct provider=openai
time="2025-11-24T11:33:54Z" level=info msg="Creating new LLM OCR provider" model=mistralai/Mistral-Small-24B-Instruct provider=openai
time="2025-11-24T11:33:54Z" level=info msg="Successfully initialized LLM OCR provider" model=mistralai/Mistral-Small-24B-Instruct provider=openai
time="2025-11-24T11:33:54Z" level=info msg="Validating OCR provider 'llm' with processing mode 'image'"
time="2025-11-24T11:33:54Z" level=info msg="✅ OCR provider and processing mode configuration is valid"
time="2025-11-24T11:33:54Z" level=info msg="Serving frontend files from embedded assets"
time="2025-11-24T11:33:54Z" level=info msg="Server started on interface :8080"
Using paperless-gpt-ocr as manual OCR tag
Using paperless-gpt-ocr-auto as auto OCR tag
time="2025-11-24T11:35:25Z" level=debug msg="Making HTTP request" method=GET url="http://webserver:8000/api/documents/?tags__name__iexact=paperless-gpt-ocr-auto&page_size=25"
time="2025-11-24T11:35:25Z" level=debug msg="Making HTTP request" method=GET url="http://webserver:8000/api/tags/"
time="2025-11-24T11:35:26Z" level=debug msg="Making HTTP request" method=GET url="http://webserver:8000/api/correspondents/?page_size=9999"
time="2025-11-24T11:35:26Z" level=debug msg="Found 1 documents with tag paperless-gpt-ocr-auto"
time="2025-11-24T11:35:26Z" level=info msg="Processing document for OCR" document_id=484
time="2025-11-24T11:35:26Z" level=info msg="Starting OCR processing" document_id=484
time="2025-11-24T11:35:26Z" level=debug msg="OCR provider does not support hOCR" document_id=484
time="2025-11-24T11:35:26Z" level=debug msg="Making HTTP request" method=GET url="http://webserver:8000/api/documents/484/download/"
time="2025-11-24T11:35:27Z" level=debug msg="Downloaded document images" document_id=484 limit_pages=0 processed_page_count=1 total_page_count=1
time="2025-11-24T11:35:27Z" level=debug msg="Processing page" document_id=484 page=1
time="2025-11-24T11:35:28Z" level=info msg="Using OpenAI image format" model=mistralai/Mistral-Small-24B-Instruct page=1 provider=openai
time="2025-11-24T11:35:28Z" level=error msg="Failed to get response from vision model" error="API returned unexpected status code: 413" model=mistralai/Mistral-Small-24B-Instruct page=1 provider=openai
time="2025-11-24T11:35:28Z" level=error msg="OCR processing failed: error performing OCR for document 484, page 1: error getting response from LLM: API returned unexpected status code: 413" document_id=484
time="2025-11-24T11:35:28Z" level=error msg="Error in background tagging: error in processAutoOcrTagDocuments: one or more errors occurred: document 484 OCR error: error performing OCR for document 484, page 1: error getting response from LLM: API returned unexpected status code: 413"
time="2025-11-24T11:35:48Z" level=debug msg="Making HTTP request" method=GET url="http://webserver:8000/api/documents/?tags__name__iexact=paperless-gpt-ocr-auto&page_size=25"
time="2025-11-24T11:35:48Z" level=debug msg="Making HTTP request" method=GET url="http://webserver:8000/api/tags/"
time="2025-11-24T11:35:48Z" level=debug msg="Making HTTP request" method=GET url="http://webserver:8000/api/correspondents/?page_size=9999"
time="2025-11-24T11:35:49Z" level=debug msg="Found 1 documents with tag paperless-gpt-ocr-auto"
time="2025-11-24T11:35:49Z" level=info msg="Processing document for OCR" document_id=484
time="2025-11-24T11:35:49Z" level=info msg="Starting OCR processing" document_id=484
time="2025-11-24T11:35:49Z" level=debug msg="OCR provider does not support hOCR" document_id=484
time="2025-11-24T11:35:49Z" level=debug msg="Making HTTP request" method=GET url="http://webserver:8000/api/documents/484/download/"
time="2025-11-24T11:35:50Z" level=debug msg="Downloaded document images" document_id=484 limit_pages=0 processed_page_count=1 total_page_count=1
time="2025-11-24T11:35:50Z" level=debug msg="Processing page" document_id=484 page=1
time="2025-11-24T11:35:51Z" level=info msg="Using OpenAI image format" model=mistralai/Mistral-Small-24B-Instruct page=1 provider=openai
time="2025-11-24T11:35:51Z" level=error msg="Failed to get response from vision model" error="API returned unexpected status code: 413" model=mistralai/Mistral-Small-24B-Instruct page=1 provider=openai
time="2025-11-24T11:35:51Z" level=error msg="OCR processing failed: error performing OCR for document 484, page 1: error getting response from LLM: API returned unexpected status code: 413" document_id=484
time="2025-11-24T11:35:51Z" level=error msg="Error in background tagging: error in processAutoOcrTagDocuments: one or more errors occurred: document 484 OCR error: error performing OCR for document 484, page 1: error getting response from LLM: API returned unexpected status code: 413"

Extract of my config:

  paperless-gpt:
    image: icereed/paperless-gpt:latest
    restart: unless-stopped
    environment:
      LLM_PROVIDER: "openai"
      OPENAI_BASE_URL: "https://openai.inference.de-txl.ionos.com/v1"
      OPENAI_API_KEY: "${OPENAI_IONOS_API_TOKEN}"
      LLM_MODEL: "openai/gpt-oss-120b"
      TOKEN_LIMIT: "5000"
      LLM_REQUESTS_PER_MINUTE: "1" # 120 Default
      OCR_PROVIDER: "llm"
      VISION_LLM_PROVIDER: "openai"
      VISION_LLM_MODEL: "mistralai/Mistral-Small-24B-Instruct"
      VISION_LLM_REQUESTS_PER_MINUTE: "1" # 120 Default
      OCR_PROCESS_MODE: "image"
      PDF_SKIP_EXISTING_OCR: "true"
      OCR_LIMIT_PAGES: "0" # Optional, default: 5. Set to 0 for no limit.
      CREATE_LOCAL_PDF: "false"
      CREATE_LOCAL_HOCR: "false"
      LOG_LEVEL: "debug"

[CyWS]

Anything I can do for debug?

[fentes]

Hey @CyWS, I'm confused about your mix between openai as VISION_LLM_PROVIDER and mistral for the used LLM Model. Did this work for you? I started to use paperpless-gpt with mistral and these settings work fine for me.

OCR_PROVIDER: "mistral_ocr"
OCR_PROCESS_MODE: "whole_pdf"
OCR_LIMIT_PAGES: "0"
MISTRAL_MODEL: "mistral-ocr-latest"
PDF_SKIP_EXISTING_OCR: "false"

[pauli2406]

@CyWS I am facing the same issue - did you manage to find a fix for this?

[b0denl0s]

same issue here.

[CyWS]

Hey @CyWS, I'm confused about your mix between openai as VISION_LLM_PROVIDER and mistral for the used LLM Model. Did this work for you? I started to use paperpless-gpt with mistral and these settings work fine for me.

It worked exactly as posted until it stopped working early November. So my config should be correct.
My understanding is that I cannot provide a custom API endpoint for OCR_PROVIDER: "mistral_ocr" and that Ionos API is only compatible with OpenAI API format.
In my use case, Mistral is only used as the model but hosted by my LLM API provider (Ionos). I have no connection or account with Mistral's own services, nor do I want to. Also pdf or whole_pdf are documented to be unavailable for OpenAI API use.

@pauli2406 @b0denl0s : Thanks for the feedback. Are you using the same API provider or another one? For me, still not fixed and I have no idea where to look or tweak anything.

[pauli2406]

Hey @CyWS, I'm confused about your mix between openai as VISION_LLM_PROVIDER and mistral for the used LLM Model. Did this work for you? I started to use paperpless-gpt with mistral and these settings work fine for me.

It worked exactly as posted until it stopped working early November. So my config should be correct.
My understanding is that I cannot provide a custom API endpoint for OCR_PROVIDER: "mistral_ocr" and that Ionos API is only compatible with OpenAI API format.
In my use case, Mistral is only used as the model but hosted by my LLM API provider (Ionos). I have no connection or account with Mistral's own services, nor do I want to. Also pdf or whole_pdf are documented to be unavailable for OpenAI API use.

@pauli2406 @b0denl0s : Thanks for the feedback. Are you using the same API provider or another one? For me, still not fixed and I have no idea where to look or tweak anything.

Hey!
Yes indeed - i also tried to use it using IONOS as a Provider - they offer a fully German hosted Version of the mistral model.
The endpoint used in this issue is allowing anyone to use it within the common open ai framework but it seems to fail here when it comes to vision tasks - I tried it with a simple curl command and had it working - when i am back home I can see if I find my snippets to replicate the behavior

[b0denl0s]

Hey @pauli2406 @CyWS @fentes

My two cents:
When using paperless-gpt with the IONOS “OpenAI-compatible” inference endpoint for Mistral Vision models, OCR/extraction fails or behaves incorrectly.

However, when I manually send a request to the same endpoint (https://openai.inference.de-txl.ionos.com/v1) using the same model (mistralai/Mistral-Small-24B-Instruct) and attach the document as a single JPEG image (~150 DPI, base64 data URL), the request succeeds and the response contains the correctly extracted text (same behavior like @pauli2406).

This strongly suggests that:

• The IONOS endpoint itself is working correctly.
• The request generated by paperless-gpt is likely not fully compatible with what the IONOS OpenAI-style API expects for vision input (message schema, image encoding, field names, etc.).

Unfortunately, even with debug logging enabled, I cannot see the actual outbound request payload that paperless-gpt sends to the Vision endpoint, so I cannot directly compare it with my working manual request.

It would be extremely helpful if the maintainer (@icereed ) could take a look at:

• The exact OpenAI-compatible JSON payload paperless-gpt generates for Vision OCR.
• Whether it uses chat/completions vs responses.
• The structure of the messages[].content array for images.
• Whether the image is sent as a proper data:image/jpeg;base64,... URL.
• ... ?

@fentes comments in other threads mention that setting:

VISION_LLM_PROVIDER=openai
VISION_LLM_MODEL: "mistralai/Mistral-Small-24B-Instruct"

looks suspicious. In this case that is intentional and correct:
IONOS exposes Mistral models through an OpenAI-compatible API, so the provider must be set to openai even though the underlying model is Mistral.

If there is a way to:

• Log the full outbound Vision request , or
• add a debug mode to dump the JSON payload, I’d be happy to test and report back.

///

My config regarding VISION OCR block:

      OCR_PROVIDER: "llm"
      OCR_PROCESS_MODE: "image"
      VISION_LLM_PROVIDER: "openai"
      VISION_LLM_MODEL: "mistralai/Mistral-Small-24B-Instruct"
      VISION_OPENAI_API_BASE: "https://openai.inference.de-txl.ionos.com/v1"
      VISION_OPENAI_API_KEY: "${IONOS_TOKEN}"

Paperless-gpt Version:

	 Version: v0.24.0
	 paperless-gpt  | Commit: ad1fdfa22b1d5958eefa6490ee092f5d18544876
	 paperless-gpt  | Build Date: 1768426089
	 paperless-gpt  | Platform: linux/amd64
	 paperless-gpt  | Go Version: go1.25.5

[CyWS]

Ok this is somehow solved for me. Ionos updated their docs and now provide dedicated OCR instructions with a dedicated OCR model (which is crazy fast, cheap with really good quality btw.). The error still occurs with the mistral model (which per doc should also be able to consume images), however not in the OCR model.

A little bit disappointing that I am not able to debug it properly and have no chance to finde the correct image size parameters. But for now, solved with the OCR model. As long as that one is producing ok output, I am good with it.

References:
The OCR model that works is lightonai/LightOnOCR-2-1B Docs
The model that throws the 413 is mistralai/Mistral-Small-24B-Instruct Docs
In case of Ionos, no other model is flagged to be capable to process image input, see feature matrix.