Yes, you can use a custom OCR model in Docling, but you must configure it according to the OCR engine you choose (EasyOCR, Tesseract, RapidOCR, or KServe v2). Each engine has its own options class where you can specify custom model or binary paths:

• EasyOCR: Set model_storage_directory in EasyOcrOptions to your custom model directory.
  ⧉

  from docling.datamodel.pipeline_options import PdfPipelineOptions, EasyOcrOptions
  pipeline_options = PdfPipelineOptions(
      do_ocr=True,
      ocr_options=EasyOcrOptions(
          lang=["en", "ru"],
          model_storage_directory="/path/to/your/easyocr/models"
      )
  )
  

• Tesseract: Set path in TesseractCliOcrOptions to your custom Tesseract binary or data path.
  ⧉

  from docling.datamodel.pipeline_options import PdfPipelineOptions, TesseractCliOcrOptions
  pipeline_options = PdfPipelineOptions(
      do_ocr=True,
      ocr_options=TesseractCliOcrOptions(
          lang=["eng", "rus"],
          path="/custom/path/to/tesseract"
      )
  )
  

• RapidOCR: Set det_model_path, cls_model_path, rec_model_path, etc., in RapidOcrOptions to your custom model files.
• KServe v2 (KserveV2OcrOptions, kind: "kserve_v2_ocr"): Connects to a remote KServe v2-compatible inference server (such as NVIDIA Triton) via gRPC or HTTP to perform OCR, without requiring local model downloads.
  ⧉

  from docling.datamodel.pipeline_options import PdfPipelineOptions, KserveV2OcrOptions
  
  pipeline_options = PdfPipelineOptions()
  pipeline_options.do_ocr = True
  pipeline_options.ocr_options = KserveV2OcrOptions(
      url="localhost:8001", # gRPC endpoint (host:port or dns:///host:port)
      model_name="ocr", # model name registered on the server
      transport="grpc", # "grpc" (default) or "http"
      lang=["english"],
      scale=2.0,
  )
  

  Key configuration fields:
  • url (required): The endpoint. For transport='http', use http(s)://host[] or plain host. For transport='grpc', use plain host or dns:///host (dns:/// enables gRPC-native DNS load balancing, useful for round_robin scheduling over headless Kubernetes services).
  • transport: "grpc" (default) or "http".
  • model_name: Model name registered on the server.
  • lang, scale: OCR language and scale parameters.
  • grpc_channel_args (gRPC only): List of extra gRPC channel arguments, e.g., [('grpc.lb_policy_name', 'round_robin')] for client-side load balancing across Kubernetes headless service endpoints when using a dns:/// URL. Default: [].
  • Optional TLS/auth settings: grpc_use_tls, grpc_metadata, headers, timeout.

You then pass pipeline_options to your DocumentConverter as usual.

If you want to use a completely custom OCR engine beyond the built-in options above, you can implement a plugin following Docling's plugin system. See this example and discussion for details.

For all engines, if your custom model files are not in the default locations, ensure your directory structure matches what the engine expects, and set the DOCLING_SERVE_ARTIFACTS_PATH environment variable to the parent directory containing all model subfolders if you want Docling to use local models for offline use. See this guide for details.

You cannot set a generic "custom OCR model path" directly in PdfPipelineOptions; you must use the engine-specific options as shown above.