Calling an endpoint when using a MinerU model catalog

After creating an endpoint by using a MinerU model catalog, you can call the endpoint’s OpenAI-compatible API through the model gateway, or use the official MinerU pipeline client to call the endpoint for document parsing.

Prerequisites

You have obtained the API URL of the endpoint. For details, see Getting the API URL.
You have created an API key that can call the endpoint. For details, see Creating an API key.

Precaution

All calls access the endpoint through the model gateway. The address format is http://<gateway>/workspace/<workspace>/endpoint/<endpoint>.
The request header must include Authorization: Bearer <api_key>, where <api_key> is the API key.
When creating an endpoint by using a MinerU model catalog, the model registry and model name are still required, but they do not take effect for the MinerU image. You can select any valid values to pass validation. The selected values do not affect the deployment result.
The default MinerU model ID is mineru-2.5. To change the model ID, configure served_model_name in the engine variables. The corresponding startup parameter is --served-model-name.
MinerU uses gpu_memory_utilization=0.9 by default. The corresponding startup parameter is --gpu-memory-utilization 0.9. To adapt to GPU splitting scenarios, adjust gpu_memory_utilization in the engine variables.

Calling the OpenAI-compatible API

Run the following command to obtain the model ID:
Terminal window
```
curl <API_URL>/v1/models \
  -H "Authorization: Bearer <api_key>"
```
If the model ID has not been adjusted through served_model_name, the default model ID is mineru-2.5.

Run the following command to initiate a single-image recognition request through the OpenAI-compatible API.

Use Base64-encoded image data.

curl <API_URL>/v1/chat/completions \
  -H "Authorization: Bearer <api_key>" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mineru-2.5",
    "messages": [
      {
        "role": "user",
        "content": [
          {
            "type": "image_url",
            "image_url": { "url": "data:image/png;base64,'"$IMG_B64"'" }
          },
          {
            "type": "text",
            "text": "OCR this image, output only the text."
          }
        ]
      }
    ]
  }'

On Linux, use IMG_B64=$(base64 -w0 <image_file>). On macOS, use IMG_B64=$(base64 -i <image_file>). Replace <image_file> with the image file to recognize.

If the model ID has been adjusted through served_model_name, replace model in the request body with the actual model ID.

Specify an image URL.

If the MinerU server can access the image URL, you can specify the image URL directly.

curl -s -X POST \
  "<API_URL>/v1/chat/completions" \
  -H "Authorization: Bearer <api_key>" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mineru-2.5",
    "messages": [
      {
        "role": "user",
        "content": [
          {
            "type": "image_url",
            "image_url": {
              "url": "https://raw.githubusercontent.com/tesseract-ocr/tessdoc/main/images/eurotext.png"
            }
          },
          {
            "type": "text",
            "text": "OCR this image, output only the text."
          }
        ]
      }
     ]
  }' | jq -r '.choices[0].message.content'

If the model ID has been adjusted through served_model_name, replace model in the request body with the actual model ID.

After the request succeeds, the text in the image is returned.

Calling the official MinerU pipeline client

Run the following command to use the official MinerU pipeline client to call the endpoint and parse an entire PDF into Markdown:

export MINERU_VL_API_KEY=<api_key>
mineru -p <pdf_file> -o <output_dir> -b vlm-http-client -u <API_URL>/

Replace <pdf_file> with the PDF file to parse, and replace <output_dir> with the output directory for the parsing result.