Before you begin

Before you start using Vector Search, you need to choose an embedding model, prepare your data, and decide what type of endpoint you'll use. This page provides some information about doing those things.

Prepare your embeddings

To use Vector Search, you need to have your embeddings ready. If you already have your embeddings, skip to Choose an endpoint.

To create your embeddings, do the following:

  1. Choose an embedding model: There are many external embedding models available, which offer different features.

    Vector Search supports dense embeddings. As Public preview features, Vector Search supports sparse embeddings and hybrid search. Hybrid search uses dense and sparse embeddings according to the weight that you specify for those embedding types.

    Depending on your use case, choose one of the following type of model:

    • Ready-to-use: If you want to semantically match text to text or text to images by the relevance of the text or image alone. This is a standard use case, so you don't need to train or tune the model. Generative AI on Vertex AI is a recommended option for this use case. Generative AI on Vertex AI uses dense embedding models.

    • Custom model for embeddings: If you want to match based on your own data or specific use case.

  2. Prepare your data: Clean and preprocess your data to ensure that it is in a format that can be used by the embedding model.

  3. Train the embedding model if you use a custom model: If you choose to use a custom embeddings model (tuning), you need to train it on your data. This can be a time-consuming process that depends on the size and complexity of your data. If you use a pretrained model from the Model Garden, then you can skip this step.

  4. Generate embeddings: After the model is trained, use it to generate embeddings for your data.

Choose an endpoint

After you have created your index, you'll deploy it to an endpoint. For more information, see Deploy and manage public index endpoints and Deploy and manage index endpoints in a VPC network. It is helpful to decide what kind of endpoint that you need before you create your index.

You can deploy your query index to one of the following:

  • Public endpoint: If you deploy to a public endpoint, you don't need to set up your network. Public networks have slightly higher latency, but are faster to set up and easier to maintain.

  • Private Endpoint: If you want to use a VPC, you must first set up networking. Vector Search supports two types of private network.

What's next

After you've generated your embeddings and decided where to deploy your index, the next step is to configure your index.