Index Size, Build Time and IVFFlat

The index can be bigger than the vectors

An HNSW index stores graph links in addition to the vectors, so it is often larger than the raw data and works best when it fits in memory (PostgreSQL's shared buffers and the OS cache). Building it on millions of rows can take a long time and a lot of memory, and heavy inserts and deletes add maintenance cost. IVFFlat is the other index type in pgvector: it clusters vectors into lists at build time (so it needs data present before you build it, and a sensible number of lists), searches the nearest probes lists, builds faster and uses less memory, but usually gives lower recall at the same speed than HNSW and may need re-building after the data distribution changes. Rule of thumb: start with HNSW unless build time or memory forces IVFFlat.

Table and index sizes, run

I ran this SQL on PostgreSQL 16 with the pgvector extension, version 0.8.6, in a Docker container. For 20,000 vectors of 32 dimensions, the raw vector data is about 2.6 MB, the table about 4 MB, and the HNSW index about 9 MB, so the index alone is several times the raw vectors. At a million vectors of 768 dimensions the same effect is gigabytes.

SELECT 'rows' AS what, count(*)::text AS value FROM big
UNION ALL SELECT 'table size (MB)', round(pg_table_size('big') / 1e6)::text
UNION ALL SELECT 'hnsw index size (MB)', round(pg_relation_size('big_hnsw') / 1e6)::text
UNION ALL SELECT 'raw vectors: 20000 x 32 x 4 bytes (MB)', round(20000 * 32 * 4 / 1e6, 1)::text;

Output:

                  what                  | value 
----------------------------------------+-------
 rows                                   | 20000
 table size (MB)                        | 4
 hnsw index size (MB)                   | 9
 raw vectors: 20000 x 32 x 4 bytes (MB) | 2.6
(4 rows)

Build after bulk loading

Creating the index after loading the data is usually much faster than inserting into an existing index.

Quick check: Which index type usually gives better recall at the same speed in pgvector?

• IVFFlat
• HNSW
• Neither; they are identical
• A B-tree