Core Models

Call CLIP and DocTR foundation models through the inference-sdk HTTP client, with sync and async methods.

InferenceHTTPClient supports core models hosted by Inference. Some of these models can be used on the Roboflow hosted inference platform (use https://serverless.roboflow.com as the URL); others can be deployed locally (usually the local server is available at http://localhost:9001).

CLIP

from inference_sdk import InferenceHTTPClient

CLIENT = InferenceHTTPClient(
    api_url="http://localhost:9001",  # or "https://serverless.roboflow.com" to use hosted serving
    api_key="ROBOFLOW_API_KEY"
)

CLIENT.get_clip_image_embeddings(inference_input="./my_image.jpg")  # single image request
CLIENT.get_clip_image_embeddings(inference_input=["./my_image.jpg", "./other_image.jpg"])  # batch image request
CLIENT.get_clip_text_embeddings(text="some")  # single text request
CLIENT.get_clip_text_embeddings(text=["some", "other"])  # other text request
CLIENT.clip_compare(
    subject="./my_image.jpg",
    prompt=["fox", "dog"],
)

The CLIENT.clip_compare(...) method allows you to compare different combinations of subject_type and prompt_type:

• (image, image) (default)

• (image, text)

• (text, image)

• (text, text)

Async methods are also available:

from inference_sdk import InferenceHTTPClient

CLIENT = InferenceHTTPClient(
    api_url="http://localhost:9001",  # or "https://serverless.roboflow.com" to use hosted serving
    api_key="ROBOFLOW_API_KEY"
)

async def see_async_method():
  await CLIENT.get_clip_image_embeddings_async(inference_input="./my_image.jpg")  # single image request
  await CLIENT.get_clip_image_embeddings_async(inference_input=["./my_image.jpg", "./other_image.jpg"])  # batch image request
  await CLIENT.get_clip_text_embeddings_async(text="some")  # single text request
  await CLIENT.get_clip_text_embeddings_async(text=["some", "other"])  # other text request
  await CLIENT.clip_compare_async(
      subject="./my_image.jpg",
      prompt=["fox", "dog"],
  )

See the CLIP model reference for details about the model itself.

DocTR

from inference_sdk import InferenceHTTPClient

CLIENT = InferenceHTTPClient(
    api_url="http://localhost:9001",  # or "https://serverless.roboflow.com" to use hosted serving
    api_key="ROBOFLOW_API_KEY"
)

CLIENT.ocr_image(inference_input="./my_image.jpg")  # single image request
CLIENT.ocr_image(inference_input=["./my_image.jpg", "./other_image.jpg"])  # batch image request

Async equivalent: CLIENT.ocr_image_async(...). See the DocTR model reference for details.

Gaze (deprecated)

CLIENT.detect_gazes(...) and CLIENT.detect_gazes_async(...) have been removed along with the MediaPipe dependency. They now short-circuit client-side and raise inference_sdk.http.errors.FeatureDeprecatedError without issuing a network call. Contact Roboflow if you require this capability.