Store vector embeddings

This page shows you how to use AlloyDB as a vector database with the vector extension that includes pgvector functions and operators. These functions and operators let you store embeddings as vector values.

Required database extension

Use the vector extension, version 0.5.0.google-1 or later, which includes pgvector functions and operators, to store generated embeddings as vector values. This is a version of pgvector that Google has extended with optimizations specific to AlloyDB.

CREATE EXTENSION IF NOT EXISTS vector;

Store generated embeddings

Ensure that you have already created a table in your AlloyDB database.

Note: If your application uses the LangChain framework and your dataset has O(100k) embeddings, we recommend that you use the AlloyDBVectorStore vector class included in the AlloyDB LangChain library to store your embeddings. For more information, see Build LLM-powered applications using LangChain.

To store vector embeddings, do the following:

1. Create a vector[] column in your table to store your embeddings:

   ALTER TABLE TABLE ADD COLUMN EMBEDDING_COLUMN vector(DIMENSIONS);
   

   Replace the following:

   • TABLE: the table name

   • EMBEDDING_COLUMN: the name of the new embedding column

   • DIMENSIONS: the number of dimensions that the model supports.

     For example, if you are using one of the text-embedding English models—for example, text-embedding-005 with Vertex AI, specify 768.

2. Copy the vectors to the vector column. The following example assumes your embeddings are available in a CSV file:

   COPY TABLE (EMBEDDING_COLUMN) FROM 'PATH_TO_VECTOR_CSV (FORMAT CSV);
   

   Replace the following:

   • PATH_TO_VECTOR_CSV: the full path of where you have stored your .csv file.

After you store the embeddings, you can use the vector extension or the alloydb_scann extension to create indexes for faster query performance.

What's next

• Create indexes and query vectors
• An example embedding workflow