Known limitations

pgvector HNSW and IVFFlat index types are not replicated by logical replication. Indexes must be created independently on each node. See Maintaining pgvector.

For very large datasets, pgvector can become prohibitively slow in terms of index creation, query throughput (queries per second), and query latency. EDB is developing edb_vectorplus, a new vector search extension that coexists with pgvector, uses the same vector type, and delivers higher query throughput and lower latency with a reduced CPU and memory footprint. Contact EDB for details.

pgvector does not support parallel index builds prior to PostgreSQL 15.