Skip to main content

Can you use ClickHouse for vector search?

· 2 min read

Yes, ClickHouse can perform vector search. The main advantages of using ClickHouse for vector search compared to using more specialized vector databases include:

  • Using ClickHouse's filtering and full-text search capabilities to refine your dataset before performing a search.
  • Performing analytics on your datasets.
  • Running a JOIN against your existing data.
  • No need to manage yet another database and complicate your infrastructure.

Here is a quick tutorial on how to use ClickHouse for vector search.

1. Create embeddings​

Your data (documents, images, or structured data) must be converted to embeddings. We recommend creating embeddings using the OpenAI Embeddings API or using the open-source Python library SentenceTransformers.

You can think of an embedding as a large array of floating-point numbers that represent your data. Check out this guide from OpenAI to learn more about embeddings.

2. Store the embeddings​

Once you have generated embeddings, you need to store them in ClickHouse. Each embedding should be stored in a separate row and can include metadata for filtering, aggregations, or analytics. Here's an example of a table that can store images with captions:

CREATE TABLE images
(
    `_file` LowCardinality(String),
    `caption` String,
    `image_embedding` Array(Float32)
)
ENGINE = MergeTree;

Let's say you want to search for pictures of dogs in your dataset. You can use a distance function like cosineDistance to take an embedding of a dog image and search for related images:

SELECT
    _file,
    caption,
    cosineDistance(
        -- An embedding of your "input" dog picture
        [0.5736801028251648, 0.2516217529773712, ...,  -0.6825592517852783],
        image_embedding
    ) AS score
FROM images
ORDER BY score ASC
LIMIT 10

This query returns the _file names and caption of the top 10 images most likely to be related to your provided dog image.

Further Reading​

To follow a more in-depth tutorial on vector search using ClickHouse, please see: