Text recognition (OCR)

maKI recognises text in images and PDF documents and returns it as Markdown, page by page. Behind it runs PaddleOCR-VL on Uni Mannheim’s GPU infrastructure; documents never leave the university network.

The endpoint follows the schema of the Mistral OCR API. It is not a chat endpoint: there is no prompt, no messages and no tokens. Document in, text out.

Endpoint	POST https://maki.uni-mannheim.de/v1/ocr
Model	paddleocr-vl
Access group	ocr (select it when requesting a key)
Input	PNG, JPEG or PDF as a data URL in the request body
Output	Markdown per page

Recognise an image

The image is passed Base64-encoded as a data URL in the document field:

MAKI_API_KEY="..."
IMAGE=$(base64 -i scan.png | tr -d '\n')

curl https://maki.uni-mannheim.de/v1/ocr \
  -H "Authorization: Bearer ${MAKI_API_KEY}" \
  -H "Content-Type: application/json" \
  -d "{
    \"model\": \"paddleocr-vl\",
    \"document\": {
      \"type\": \"image_url\",
      \"image_url\": \"data:image/png;base64,${IMAGE}\"
    }
  }"

Response:

{
  "object": "ocr",
  "model": "paddleocr-vl",
  "pages": [
    {
      "index": 0,
      "markdown": "Invoice no. 2026-0815\n\nUniversität Mannheim, Schloss\n\nAmount: 1,234.56 EUR",
      "images": [],
      "dimensions": { "dpi": null, "height": null, "width": null }
    }
  ],
  "usage_info": { "pages_processed": 1 }
}

Recognise a PDF

For PDFs the document type is document_url. Every page of the PDF becomes one entry in pages, in order, with index starting at 0:

PDF=$(base64 -i minutes.pdf | tr -d '\n')

curl https://maki.uni-mannheim.de/v1/ocr \
  -H "Authorization: Bearer ${MAKI_API_KEY}" \
  -H "Content-Type: application/json" \
  -d "{
    \"model\": \"paddleocr-vl\",
    \"document\": {
      \"type\": \"document_url\",
      \"document_url\": \"data:application/pdf;base64,${PDF}\"
    }
  }"

Python

The endpoint is not part of the OpenAI SDK; a plain HTTP call is enough:

import base64
import requests

MAKI_API_KEY = "..."

with open("minutes.pdf", "rb") as f:
    data = base64.b64encode(f.read()).decode()

response = requests.post(
    "https://maki.uni-mannheim.de/v1/ocr",
    headers={"Authorization": f"Bearer {MAKI_API_KEY}"},
    json={
        "model": "paddleocr-vl",
        "document": {
            "type": "document_url",
            "document_url": f"data:application/pdf;base64,{data}",
        },
    },
    timeout=120,
)
response.raise_for_status()

for page in response.json()["pages"]:
    print(f"--- page {page['index'] + 1} ---")
    print(page["markdown"])

For images: "type": "image_url", "image_url": "data:image/png;base64,...".

Notes

• Data URLs only. An https:// address as image_url or document_url is currently not fetched and rejected with HTTP 400. Download the document yourself and pass it Base64-encoded.
• All pages. The pages field for page selection is still ignored, as is include_image_base64; images stays empty. The response always contains every page of the document.
• Latency. One page takes about one to two seconds. Documents are processed in blocks of twelve pages; split very long PDFs beforehand, since a request is cut off after 75 seconds.
• Size. Base64 inflates the file by a third. Scale images to a sensible resolution before sending (text clearly legible, roughly 150 to 300 dpi).
• No prompt. To process the recognised text further (summarise, extract fields), send the Markdown to a chat model in a second step, see Structured output.