About vector search

Here’s a closer look at how Serverless Cassandra with Vector Search performs and why.

CassIO

For typical generative artificial intelligence (AI) or other machine learning workloads, CassIO abstracts away the details of accessing the Cassandra database. CassIO offers a low-boilerplate, ready-to-use set of tools for seamless integration of Cassandra in most AI-oriented applications. For more, see CassIO.

Preprocessing embedding vectors

Depending on the nature of the data, to get the best from the vector search, some form of preprocessing might be necessary before writing to the database. Two common preprocessing methods are normalizing and standardizing.

Method Definition Features

Method	Definition	Features
Normalizing (Not required for all vector examples)	Scale data to a length of one by dividing each element in a vector by the vector’s length (also known as its `Euclidean` norm or L2 norm). If embeddings are NOT normalized, `dot product` silently returns meaningless query results. Therefore, `dot product` is not set as the default similarity function in Serverless Cassandra with Vector Search.	Eliminates the impact of vector scale and focuses on their direction. Makes high-dimensional data consistent `Dot product` algorithm becomes faster
Standardizing	Shift and scale data for a mean of zero and a standard deviation of one	Aims to give them properties of a standard normal Gaussian distribution. Crucial in ML models that use distances between points (ensures all features contribute equally to distance calculations, prevents larger scale features from dominating) Maintains a balanced scale among features

Normalizing

(Not required for all vector examples)

Scale data to a length of one by dividing each element in a vector by the vector’s length (also known as its Euclidean norm or L2 norm).

If embeddings are NOT normalized, dot product silently returns meaningless query results. Therefore, dot product is not set as the default similarity function in Serverless Cassandra with Vector Search.

Eliminates the impact of vector scale and focuses on their direction.
Makes high-dimensional data consistent
Dot product algorithm becomes faster

Standardizing

Shift and scale data for a mean of zero and a standard deviation of one

Aims to give them properties of a standard normal Gaussian distribution.
Crucial in ML models that use distances between points (ensures all features contribute equally to distance calculations, prevents larger scale features from dominating)
Maintains a balanced scale among features

When you use OpenAI, PaLM, or Simsce to generate your embeddings, they are normalized by default. If you use a different library, normalize your vectors and set the similarity function to dot product. See how to set the similarity function in Query vector data with CQL.

Similarity metric

Vector search requires appropriate similarity metrics such as cosine, Euclidean, or Jaccard. The data type and desired search behavior determines the metric. Our search supports cosine (default), dot product, and Euclidean similarities.

DataStax recommends using dot product on normalized embeddings for most applications because of its superior speed.

Scalability and performance

Scalability is a critical consideration as your dataset expands. Vector search algorithms should be designed to handle large-scale datasets efficiently. The database distributes data and accesses that data with parallel processing, enhancing performance.

Evaluation and iteration

Iteratively refining the vector representations, similarity metrics, indexing techniques, or preprocessing steps can lead to better search performance and user satisfaction. Continuously evaluate and iterate the data to refine search results against known truth and user feedback.

Integrations with vector search

Third-party integrations connect your Serverless Cassandra with Vector Search to various Large Language Model (LLM) frameworks. Use integrations to streamline your vector-based similarity searches and to aid in developing LLM-powered applications.

For the full list of third-party integrations, see Astra DB integrations.