Skip to main content

Locally Hosted CLIP Embeddings + Weaviate

Weaviate's integration with the Hugging Face Transformers library allows you to access their CLIP models' capabilities directly from Weaviate.

Configure a Weaviate vector index to use the CLIP integration, and configure the Weaviate instance with a model image, and Weaviate will generate embeddings for various operations using the specified model in the CLIP inference container. This feature is called the vectorizer.

At import time, Weaviate generates multimodal object embeddings and saves them into the index. For vector and hybrid search operations, Weaviate converts queries of one or more modalities into embeddings.

Embedding integration illustration

Requirements​

Weaviate configuration​

Your Weaviate instance must be configured with the CLIP vectorizer integration (multi2vec-clip) module.

For Weaviate Cloud (WCD) users

This integration is not available for Weaviate Cloud (WCD) serverless instances, as it requires spinning up a container with the Hugging Face model.

Enable the integration module​

Configure the integration​

To use this integration, configure the container image of the CLIP model and the inference endpoint of the containerized model.

The following example shows how to configure the CLIP integration in Weaviate:

Docker Option 1: Use a pre-configured docker-compose.yml file​

Follow the instructions on the Weaviate Docker installation configurator to download a pre-configured docker-compose.yml file with a selected model


Docker Option 2: Add the configuration manually​

Alternatively, add the configuration to the docker-compose.yml file manually as in the example below.

services:
  weaviate:
    # Other Weaviate configuration
    environment:
      CLIP_INFERENCE_API: http://multi2vec-clip:8080  # Set the inference API endpoint
  multi2vec-clip:  # Set the name of the inference container
    image: cr.weaviate.io/semitechnologies/multi2vec-clip:sentence-transformers-clip-ViT-B-32-multilingual-v1
    environment:
      ENABLE_CUDA: 0  # Set to 1 to enable
  • CLIP_INFERENCE_API environment variable sets the inference API endpoint
  • multi2vec-clip is the name of the inference container
  • image is the container image
  • ENABLE_CUDA environment variable enables GPU usage

Set image from a list of available models to specify a particular model to be used.

Credentials​

As this integration runs a local container with the CLIP model, no additional credentials (e.g. API key) are required. Connect to Weaviate as usual, such as in the examples below.

import weaviate

client = weaviate.connect_to_local()

# Work with Weaviate

client.close()

Configure the vectorizer​

Configure a Weaviate index as follows to use a CLIP embedding model:

from weaviate.classes.config import Configure

client.collections.create(
    "DemoCollection",
    properties=[
        Property(name="title", data_type=DataType.TEXT),
        Property(name="poster", data_type=DataType.BLOB),
    ],
    vectorizer_config=[
        Configure.NamedVectors.multi2vec_clip(
            name="title_vector",
            # Define the fields to be used for the vectorization - using image_fields, text_fields, video_fields
            image_fields=[
                Multi2VecField(name="poster", weight=0.9)
            ],
            text_fields=[
                Multi2VecField(name="title", weight=0.1)
            ]
        )
    ],
    # Additional parameters not shown
)
Chose a container image to select a model

To chose a model, select the container image that hosts it.

Vectorization behavior

Weaviate follows the collection configuration and a set of predetermined rules to vectorize objects.


Unless specified otherwise in the collection definition, the default behavior is to:


  • Only vectorize properties that use the text or text[] data type (unless skipped)
  • Sort properties in alphabetical (a-z) order before concatenating values
  • If vectorizePropertyName is true (false by default) prepend the property name to each property value
  • Join the (prepended) property values with spaces
  • Prepend the class name (unless vectorizeClassName is false)
  • Convert the produced string to lowercase

Vectorizer parameters​

Inference URL parameters​

Optionally, if your stack includes multiple inference containers, specify the inference container(s) to use with a collection.

If no parameters are specified, the default inference URL from the Weaviate configuration is used.

Specify inferenceUrl for a single inference container.

from weaviate.classes.config import Configure

client.collections.create(
    "DemoCollection",
    properties=[
        Property(name="title", data_type=DataType.TEXT),
        Property(name="poster", data_type=DataType.BLOB),
    ],
    vectorizer_config=[
        Configure.NamedVectors.multi2vec_clip(
            name="title_vector",
            # Define the fields to be used for the vectorization - using image_fields, text_fields, video_fields
            image_fields=[
                Multi2VecField(name="poster", weight=0.9)
            ],
            text_fields=[
                Multi2VecField(name="title", weight=0.1)
            ],
            # inference_url="<custom_clip_url>"
        )
    ],
    # Additional parameters not shown
)

Data import​

After configuring the vectorizer, import data into Weaviate. Weaviate generates embeddings for the objects using the specified model.

collection = client.collections.get("DemoCollection")

with collection.batch.dynamic() as batch:
    for src_obj in source_objects:
        poster_b64 = url_to_base64(src_obj["poster_path"])
        weaviate_obj = {
            "title": src_obj["title"],
            "description": src_obj["description"],
            "poster": poster_b64  # Add the image in base64 encoding
        }

        # The model provider integration will automatically vectorize the object
        batch.add_object(
            properties=weaviate_obj,
            # vector=vector  # Optionally provide a pre-obtained vector
        )
Re-use existing vectors

If you already have a compatible model vector available, you can provide it directly to Weaviate. This can be useful if you have already generated embeddings using the same model and want to use them in Weaviate, such as when migrating data from another system.

Searches​

Once the vectorizer is configured, Weaviate will perform vector and hybrid search operations using the specified CLIP model.

Embedding integration at search illustration

Vector (near text) search​

When you perform a vector search, Weaviate converts the text query into an embedding using the specified model and returns the most similar objects from the database.

The query below returns the n most similar objects from the database, set by limit.

collection = client.collections.get("DemoCollection")

response = collection.query.near_text(
    query="A holiday film",  # The model provider integration will automatically vectorize the query
    limit=2
)

for obj in response.objects:
    print(obj.properties["title"])
What is a hybrid search?

A hybrid search performs a vector search and a keyword (BM25) search, before combining the results to return the best matching objects from the database.

When you perform a hybrid search, Weaviate converts the text query into an embedding using the specified model and returns the best scoring objects from the database.

The query below returns the n best scoring objects from the database, set by limit.

collection = client.collections.get("DemoCollection")

response = collection.query.hybrid(
    query="A holiday film",  # The model provider integration will automatically vectorize the query
    limit=2
)

for obj in response.objects:
    print(obj.properties["title"])

When you perform a media search such as a near image search, Weaviate converts the query into an embedding using the specified model and returns the most similar objects from the database.

To perform a near media search such as near image search, convert the media query into a base64 string and pass it to the search query.

The query below returns the n most similar objects to the input image from the database, set by limit.

def url_to_base64(url):
    import requests
    import base64

    image_response = requests.get(url)
    content = image_response.content
    return base64.b64encode(content).decode("utf-8")


collection = client.collections.get("DemoCollection")

query_b64 = url_to_base64(src_img_path)

response = collection.query.near_image(
    near_image=query_b64,
    limit=2,
    return_properties=["title", "release_date", "tmdb_id", "poster"]  # To include the poster property in the response (`blob` properties are not returned by default)
)

for obj in response.objects:
    print(obj.properties["title"])

References​

Available models​

Lists of pre-built Docker images for this integration are below.

Model NameImage NameNotes
sentence-transformers-clip-ViT-B-32cr.weaviate.io/semitechnologies/multi2vec-clip:sentence-transformers-clip-ViT-B-32Texts must be in English. (English, 768d)
sentence-transformers-clip-ViT-B-32-multilingual-v1cr.weaviate.io/semitechnologies/multi2vec-clip:sentence-transformers-clip-ViT-B-32-multilingual-v1Supports a wide variety of languages for text. See sbert.net for details. (Multilingual, 768d)
openai-clip-vit-base-patch16cr.weaviate.io/semitechnologies/multi2vec-clip:openai-clip-vit-base-patch16The base model uses a ViT-B/16 Transformer architecture as an image encoder and uses a masked self-attention Transformer as a text encoder.
ViT-B-16-laion2b_s34b_b88kcr.weaviate.io/semitechnologies/multi2vec-clip:ViT-B-16-laion2b_s34b_b88kThe base model uses a ViT-B/16 Transformer architecture as an image encoder trained with LAION-2B dataset using OpenCLIP.
ViT-B-32-quickgelu-laion400m_e32cr.weaviate.io/semitechnologies/multi2vec-clip:ViT-B-32-quickgelu-laion400m_e32The base model uses a ViT-B/32 Transformer architecture as an image encoder trained with LAION-400M dataset using OpenCLIP.
xlm-roberta-base-ViT-B-32-laion5b_s13b_b90kcr.weaviate.io/semitechnologies/multi2vec-clip:xlm-roberta-base-ViT-B-32-laion5b_s13b_b90kUses ViT-B/32 xlm roberta base model trained with the LAION-5B dataset using OpenCLIP.

We add new model support over time. For a complete list of available models, see the Docker Hub tags for the multi2vec-clip container.

Advanced configuration​

Run a separate inference container​

As an alternative, you can run the inference container independently from Weaviate. To do so, follow these steps:

  • Enable multi2vec-clip and omit multi2vec-clip container parameters in your Weaviate configuration
  • Run the inference container separately, e.g. using Docker, and
  • Use CLIP_INFERENCE_API or inferenceUrl to set the URL of the inference container.

For example, run the container with Docker:

docker run -itp "8000:8080" semitechnologies/multi2vec-clip:sentence-transformers-clip-ViT-B-32-multilingual-v1

Then, set CLIP_INFERENCE_API="http://localhost:8000". If Weaviate is part of the same Docker network, as a part of the same docker-compose.yml file, you can use the Docker networking/DNS, such as CLIP_INFERENCE_API=http://multi2vec-clip:8080.

Further resources​

Other integrations​

Code examples​

Once the integrations are configured at the collection, the data management and search operations in Weaviate work identically to any other collection. See the following model-agnostic examples:

  • The how-to: manage data guides show how to perform data operations (i.e. create, update, delete).
  • The how-to: search guides show how to perform search operations (i.e. vector, keyword, hybrid) as well as retrieval augmented generation.

Model licenses​

Each of the compatible models has its own license. For detailed information, review the license for the model you are using in the Hugging Face Model Hub.

It is your responsibility to evaluate whether the terms of its license(s), if any, are appropriate for your intended use.

External resources​

Questions and feedback​

If you have any questions or feedback, let us know in the user forum.