Transactions and Types: vector, halfvec

Vectors get database guarantees

Since vectors are ordinary rows, you get transactions: update a document's text and its embedding atomically, roll back a failed batch, delete rows and have indexes follow, and take consistent backups with the tools you already use (pg_dump, base backups, point-in-time recovery). pgvector also offers compact types: halfvec stores 16-bit floats (half the size of vector), and there are sparse and bit vector types for other uses; check the extension's documentation for the limits in your version, such as the maximum dimensions you can index. Smaller types reduce memory and disk, usually with little recall loss, but measure on your data.

A rolled-back delete, run

I ran this SQL on PostgreSQL 16 with the pgvector extension, version 0.8.6, in a Docker container. Inside the transaction the tenant beta row is gone (4 rows); after ROLLBACK all 5 rows are back. The vector column takes part in the transaction like any other.

BEGIN;
DELETE FROM docs WHERE tenant = 'beta';
SELECT count(*) AS inside_transaction FROM docs;
ROLLBACK;
SELECT count(*) AS after_rollback FROM docs;

Output:

 inside_transaction 
--------------------
                  4
(1 row)

 after_rollback 
----------------
              5
(1 row)

halfvec is smaller, run

I ran this SQL on PostgreSQL 16 with the pgvector extension, version 0.8.6, in a Docker container. A 4-dimension vector value takes 24 bytes in this measurement and halfvec takes 16, because 16-bit floats are half the size of 32-bit ones (the fixed header is the same).

SELECT pg_column_size('[0.1,0.2,0.3,0.4]'::vector(4))  AS vector_bytes,
       pg_column_size('[0.1,0.2,0.3,0.4]'::halfvec(4)) AS halfvec_bytes;

Output:

 vector_bytes | halfvec_bytes 
--------------+---------------
           24 |            16
(1 row)

Quick check: What benefit do transactions give vector data?

• Higher recall automatically
• Faster embeddings
• Atomic updates of text and embedding, and rollback
• No need for backups