Skip to main content

Create a collection

To use named vectors, your collection be configured with named vector definitions.

Code​

This example creates a collection for the movie data, including multiple named vector definitions:

import weaviate

import weaviate.classes.config as wc


# Instantiate your client (not shown). e.g.:
# headers = {"X-OpenAI-Api-Key": os.getenv("OPENAI_APIKEY")}  # Replace with your OpenAI API key
# client = weaviate.connect_to_local(headers=headers)

client.collections.create(
    name="MovieNVDemo",  # The name of the collection ('NV' for named vectors)
    properties=[
        wc.Property(name="title", data_type=wc.DataType.TEXT),
        wc.Property(name="overview", data_type=wc.DataType.TEXT),
        wc.Property(name="vote_average", data_type=wc.DataType.NUMBER),
        wc.Property(name="genre_ids", data_type=wc.DataType.INT_ARRAY),
        wc.Property(name="release_date", data_type=wc.DataType.DATE),
        wc.Property(name="tmdb_id", data_type=wc.DataType.INT),
        wc.Property(name="poster", data_type=wc.DataType.BLOB),
    ],
    # Define & configure the vectorizer module
    vectorizer_config=[
        # Vectorize the movie title
        wc.Configure.NamedVectors.text2vec_openai(
            name="title", source_properties=["title"]
        ),
        # Vectorize the movie overview (summary)
        wc.Configure.NamedVectors.text2vec_openai(
            name="overview", source_properties=["overview"]
        ),
        # Vectorize the movie poster & title
        wc.Configure.NamedVectors.multi2vec_clip(
            name="poster_title",
            image_fields=[
                wc.Multi2VecField(name="poster", weight=0.9)
            ],  # 90% of the vector is from the poster
            text_fields=[
                wc.Multi2VecField(name="title", weight=0.1)
            ],  # 10% of the vector is from the title
        ),
    ],
    # Define the generative module
    generative_config=wc.Configure.Generative.openai(),
)

client.close()

Explain the code​

The key difference here is the use of NamedVectors class to define vectorizer configurations. Let's review the code in further detail:

Revision

This code builds on the multimodal example. Review that example for further explanations.

Named vector configuration​

This definition allows each object to be represented by three vectors, named title, overview and poster_title.

        # Vectorize the movie title
        wc.Configure.NamedVectors.text2vec_openai(
            name="title", source_properties=["title"]
        ),
        # Vectorize the movie overview (summary)
        wc.Configure.NamedVectors.text2vec_openai(
            name="overview", source_properties=["overview"]
        ),
        # Vectorize the movie poster & title
        wc.Configure.NamedVectors.multi2vec_clip(
            name="poster_title",
            image_fields=[
                wc.Multi2VecField(name="poster", weight=0.9)
            ],  # 90% of the vector is from the poster
            text_fields=[
                wc.Multi2VecField(name="title", weight=0.1)
            ],  # 10% of the vector is from the title
        ),

title​

This vector representation is generated from the title property (source_properties). The text2vec-openai module is used for vectorization.

You could use this to search for movies by similarities to their titles.

overview​

This vector representation is based on the overview property. As such, you could use this to search for movies by similarities to their plot or key ideas.

poster_title​

This vector representation is generated from a combination of the title and poster properties. The multi2vec-clip module is used for vectorization.

Note that the majority of the vector weight is given to the poster property (90%), and the rest to the title property (10%). This means that the vector representation will be more influenced by the poster than the title.

As this uses a multimodal vectorizer, you could use this to search for movies using any image, or text, by their similarity to the title or poster.

Questions and feedback​

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