Issue with Azure Mistral OCR Deployment – Always Returning null

Hi @Francisco dos Santos,

Thank you for confirming the troubleshooting steps you've already taken and for providing the detailed information.

It does appear that the behavior you're encountering is not expected. Please try below steps:

Although the documentation suggests using image_url, some OCR models deployed via Azure AI Studio may expect file inputs differently when working with Base64 content. Instead of image_url, could you try sending the payload using the file field directly with the raw Base64-encoded PDF? Some models treat image_url strictly as a URL to an image hosted externally rather than Base64 content.

If supported in the model’s endpoint, sending the document as a direct file upload (using the file field) may produce a valid output instead of using image_url. This distinction sometimes causes silent failures without error messages.

Since this is a Mistral OCR model integrated through Azure AI Foundry or Azure AI Studio, model-specific guidance might differ slightly from standard public Mistral API behavior. I recommend reviewing any model-specific guidelines published in the Azure AI Model Catalog for Mistral OCR (if available).

To rule out any deployment-specific issue, I can suggest either redeploying the model or creating a new deployment instance and testing the same document. Sometimes initial deployments can encounter silent configuration issues.

Please let me know if you have further queries.

Thank you!

try this

class MistralOCR:
    def __init__(self): 
        self.api_key = ""
        self.endpoint = ""
        self.model = "mistral-ocr-2503"

    def encode_pdf_as_base64(self, pdf_stream: BytesIO) -> str:
        """Encode PDF content as base64 string"""
        try:
            pdf_stream.seek(0)
            pdf_bytes = pdf_stream.read()
            base64_encoded = base64.b64encode(pdf_bytes).decode('utf-8')
            return base64_encoded
        except Exception as e:
            print(f"Error encoding PDF as base64: {e}")
            raise

    def _create_payload(self, pdf_base64: str) -> dict[str, Any]:
        return {
            "model": self.model,
            "document": {
                "document_url": f"data:application/pdf;base64,{pdf_base64}"

            },
            "include_image_base64": True
        }

    def generate_response(self, pdf_base64: str) -> Any:
        headers = {
            "Content-Type": "application/json",
            "Authorization": self.api_key,
        }

        payload = self._create_payload(pdf_base64)
        response = requests.post(self.endpoint, headers=headers, data=json.dumps(payload))
        
        response.raise_for_status()
        return response.json()

You have to put document_url and not image_url. The documentation is wrong.

Hi @Francisco dos Santos,

Could you please try this

from azure.ai.inference import InferenceClient
import os

endpoint = os.environ.get("AZUREAI_ENDPOINT")
api_key = os.environ.get("AZUREAI_API_KEY")
client = InferenceClient(endpoint=endpoint, api_key=api_key)

response = client.chat_completions.create(
    model="mistral-ocr-2503",
    document={"type": "document_url", "document_url": "your_document_url"},
    include_image_base64=True
)
print(response)

The key factor might be the document type— you could try something straightforward like making a REST call using curl.

curl -X POST $AZUREAI_ENDPOINT/v1/models/mistral-ocr-2503/infer \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $AZURE_API_KEY" \
  -d '{"document": {"type": "document_url", "document_url": "your_document_url"}, "include_image_base64": true}'

Make sure the URL you're testing with isn’t protected use a publicly accessible image instead. If every response returns a 200 status with no data, and you've only tested with files stored in secure locations, try using an image from your organization's public website or a direct link to an image with text from Bing (e.g., an example Mistral image from GitHub or Bing Copilot). This helps confirm whether the issue lies with the input or the actual endpoint configuration/deployment.

Please refer this https://docs.aimlapi.com/api-references/vision-models/ocr-optical-character-recognition/mistral-ai/mistral-ocr-latest

Thank you!

Hi @Francisco dos Santos,

We haven’t heard from you on the last response and was just checking back to see if you have a resolution yet. In case if you have any resolution please do share that same with the community as it can be helpful to others. Otherwise, will respond with more details and we will try to help.

Thank you!

Hello I disagree with the solution proposed. It doesn't work correctly - almost half time. Using pdf full content or only one image content, there's always a reliability problem.

I also disagree with the answer. I tried type with both document_url and image_url and I get null, but weirdly, only on the first request attempt at parsing the document. If I retry a second request in the same minute with the same document, it parses it correctly.

If I upload another document, it returns null again on the first request attempt, but returns the correct results on the next requests.