Vector set embeddings

A Redis [vector set]({{< relref "/develop/data-types/vector-sets" >}}) lets you store a set of unique keys, each with its own associated vector. You can then retrieve keys from the set according to the similarity between their stored vectors and a query vector that you specify.

You can use vector sets to store any type of numeric vector but they are particularly optimized to work with text embedding vectors (see [Redis for AI]({{< relref "/develop/ai" >}}) to learn more about text embeddings). The example below shows how to use the sentence-transformers library to generate vector embeddings and then store and retrieve them using a vector set with redis-py.

Initialize

Start by installing the preview version of redis-py with the following command:

pip install redis==6.0.0b2

Also, install sentence-transformers:

pip install sentence-transformers

In a new Python file, import the required classes:

{{< clients-example set="home_vecsets" step="import" lang_filter="Python" description="Foundational: Import required libraries for vector sets, embeddings, and Redis operations" difficulty="beginner" >}} {{< /clients-example >}}

The first of these imports is the SentenceTransformer class, which generates an embedding from a section of text. This example uses an instance of SentenceTransformer with the all-MiniLM-L6-v2 model for the embeddings. This model generates vectors with 384 dimensions, regardless of the length of the input text, but note that the input is truncated to 256 tokens (see Word piece tokenization at the Hugging Face docs to learn more about the way tokens are related to the original text).

{{< clients-example set="home_vecsets" step="model" lang_filter="Python" description="Foundational: Initialize a SentenceTransformer model to generate vector embeddings from text" difficulty="beginner" >}} {{< /clients-example >}}

Create the data

The example data is contained a dictionary with some brief descriptions of famous people:

{{< clients-example set="home_vecsets" step="data" lang_filter="Python" description="Foundational: Define sample data with text descriptions for vector embedding and storage" difficulty="beginner" >}} {{< /clients-example >}}

Add the data to a vector set

The next step is to connect to Redis and add the data to a new vector set.

The code below uses the dictionary's items() view to iterate through all the key-value pairs and add corresponding elements to a vector set called famousPeople.

Use the encode() method of SentenceTransformer to generate the embedding as an array of float32 values. The tobytes() method converts the array to a byte string that you can pass to the [vadd()]({{< relref "/commands/vadd" >}}) command to set the embedding. Note that vadd() can also accept a list of float values to set the vector, but the byte string format is more compact and saves a little transmission time. If you later use [vemb()]({{< relref "/commands/vemb" >}}) to retrieve the embedding, it will return the vector as an array rather than the original byte string (note that this is different from the behavior of byte strings in [hash vector indexing]({{< relref "/develop/ai/search-and-query/vectors" >}})).

The call to vadd() also adds the born and died values from the original dictionary as attribute data. You can access this during a query or by using the [vgetattr()]({{< relref "/commands/vgetattr" >}}) method.

{{< clients-example set="home_vecsets" step="add_data" lang_filter="Python" description="Foundational: Add vector embeddings and attributes to a vector set using VADD command" difficulty="beginner" >}} {{< /clients-example >}}

Query the vector set

You can now query the data in the set. The basic approach is to use the encode() method to generate another embedding vector for the query text. (This is the same method used to add the elements to the set.) Then, pass the query vector to [vsim()]({{< relref "/commands/vsim" >}}) to return elements of the set, ranked in order of similarity to the query.

Start with a simple query for "actors":

{{< clients-example set="home_vecsets" step="basic_query" lang_filter="Python" description="Vector similarity search: Find semantically similar items in a vector set using VSIM command" difficulty="intermediate" >}} {{< /clients-example >}}

This returns the following list of elements (formatted slightly for clarity):

'actors': ['Masako Natsume', 'Chaim Topol', 'Linus Pauling',
'Marie Fredriksson', 'Maryam Mirzakhani', 'Marie Curie',
'Freddie Mercury', 'Paul Erdos']

The first two people in the list are the two actors, as expected, but none of the people from Linus Pauling onward was especially well-known for acting (and there certainly isn't any information about that in the short description text). As it stands, the search attempts to rank all the elements in the set, based on the information contained in the embedding model. You can use the count parameter of vsim() to limit the list of elements to just the most relevant few items:

{{< clients-example set="home_vecsets" step="limited_query" lang_filter="Python" description="Vector similarity search with limits: Restrict results to the top K most similar items using the count parameter" difficulty="intermediate" >}} {{< /clients-example >}}

The reason for using text embeddings rather than simple text search is that the embeddings represent semantic information. This allows a query to find elements with a similar meaning even if the text is different. For example, the word "entertainer" doesn't appear in any of the descriptions but if you use it as a query, the actors and musicians are ranked highest in the results list:

{{< clients-example set="home_vecsets" step="entertainer_query" lang_filter="Python" description="Semantic search: Leverage text embeddings to find semantically similar items even when exact keywords don't match" difficulty="intermediate" >}} {{< /clients-example >}}

Similarly, if you use "science" as a query, you get the following results:

'science': ['Marie Curie', 'Linus Pauling', 'Maryam Mirzakhani',
'Paul Erdos', 'Marie Fredriksson', 'Freddie Mercury', 'Masako Natsume',
'Chaim Topol']

The scientists are ranked highest but they are then followed by the mathematicians. This seems reasonable given the connection between mathematics and science.

You can also use [filter expressions]({{< relref "/develop/data-types/vector-sets/filtered-search" >}}) with vsim() to restrict the search further. For example, repeat the "science" query, but this time limit the results to people who died before the year 2000:

{{< clients-example set="home_vecsets" step="filtered_query" lang_filter="Python" description="Filtered vector search: Combine vector similarity with attribute filters to narrow results based on metadata conditions" difficulty="advanced" >}} {{< /clients-example >}}

Note that the boolean filter expression is applied to items in the list before the vector distance calculation is performed. Items that don't pass the filter test are removed from the results completely, rather than just reduced in rank. This can help to improve the performance of the search because there is no need to calculate the vector distance for elements that have already been filtered out of the search.

More information

See the [vector sets]({{< relref "/develop/data-types/vector-sets" >}}) docs for more information and code examples. See the [Redis for AI]({{< relref "/develop/ai" >}}) section for more details about text embeddings and other AI techniques you can use with Redis.

You may also be interested in [vector search]({{< relref "/develop/clients/redis-py/vecsearch" >}}). This is a feature of [Redis Search]({{< relref "/develop/ai/search-and-query" >}}) that lets you retrieve [JSON]({{< relref "/develop/data-types/json" >}}) and [hash]({{< relref "/develop/data-types/hashes" >}}) documents based on vector data stored in their fields.