Skip to content

Natural Language Search Setup#

Kolena supports natural language and similar image search across image data registered to the platform. Users may set up this functionality by extracting and uploading the corresponding search embeddings using a Kolena provided package. In this document, we will go over main components of the example and steps you need to take to tailor it for your application.


The kolena repository contains a runnable example for embeddings extraction and upload. This builds off the data uploaded in the semantic_segmentation example dataset, and is best run after this data has been uploaded to your Kolena environment.

Uploading embeddings to Kolena can be done in three simple steps:

  • Step 1: installing dependency package
  • Step 2: loading dataset and model to run embedding extraction
  • Step 3: loading images for input to extraction library
  • Step 4: extracting and uploading search embeddings

Let's take a look at each step with example code snippets.

Step 1: Install kolena-embeddings Package#

The package can be installed via pip or poetry and requires use of your kolena token which can be created on the Developer page.

We first retrieve and set our KOLENA_TOKEN environment variable. This is used by the uploader for authentication against your Kolena instance.

export KOLENA_TOKEN="********"

Run the following command, making sure to replace with the token retrieved from the developer page:

pip install --extra-index-url="https://<KOLENA_TOKEN>" kolena-embeddings

Configure an additional poetry source:

poetry source add --priority=supplemental kolena-embeddings ""

Then run the following command, making sure to replace with the token retrieved from the developer page:

poetry config http-basic.kolena-embeddings <KOLENA_TOKEN> ""

This package provides the kembed.util.extract_embeddings method that generates embeddings as a numpy array for a given PIL.Image.Image object.

Step 2: Load Dataset and Model#

Before extracting embeddings on a dataset, we need to load the dataset. The dataset seeded in the semantic_segmentation example contains image assets referenced by the locator column, and we load the dataset in to a dataframe.

The embedding model and its key are obtained via the load_embedding_model() method.

df_dataset = download_dataset("coco-stuff-10k")
model, model_key = load_embedding_model()

Step 3: Load Images for Extraction#

In order to extract embeddings on image data, we must load our image files into a PIL.Image.Image object. In this section, we will load these images from an S3 bucket. For other cloud storage services, please refer to your cloud storage's API docs.

s3 = boto3.client("s3")

def load_image_from_accessor(accessor: str) -> Image:
    bucket_name, *parts = accessor[5:].split("/")
    file_stream = boto3.resource("s3").Bucket(bucket_name).Object("/".join(parts)).get()["Body"]

def iter_image_paths(image_accessors: List[str]) -> Iterator[Tuple[str, Image.Image]]:
    for locator in image_accessors:
        image = load_image_from_accessor(locator)
        yield locator, image


When processing large scales of images, we recommend using an Iterator to limit the number of images loaded into memory at once.

Step 4: Extract and Upload Embeddings#

Once embeddings are extracted for each locator on the dataset, we create a dataframe with embedding and locator columns, and use the upload_dataset_embeddings method to upload the embeddings.

The dataframe uploaded is required to contain the ID columns of the dataset in order to match against the datapoints in the dataset. In this example, the ID column of the dataset is locator.

def extract_image_embeddings(
    model: StudioModel,
    locators_and_filepaths: List[Tuple[str, Optional[str]]],
    batch_size: int = 50,
) -> List[Tuple[str, np.ndarray]]:
    Extract a list of search embeddings corresponding to sample locators.

locator_and_image_iterator = iter_image_paths(locators)
locator_and_embeddings = extract_image_embeddings(model, locator_and_image_iterator)

df_embeddings = pd.DataFrame(locator_and_embeddings, columns=["locator", "embedding"])
upload_dataset_embeddings(dataset_name, model_key, df_embeddings)

Once the upload completes, we can now visit Datasets, open the dataset and navigate to the Studio tab to search by natural language or similar images over the corresponding image data.


In this tutorial, we learned how to extract and upload vector embeddings over your image data.


Can I share embeddings with Kolena even if I do not share the underlying images?


Embeddings extraction is a unidirectional mapping, and used only for natural language search and similarity comparisons. Uploading these embeddings to Kolena does not allow for any reconstruction of these images, nor does it involve sharing these images with Kolena.