The vector column type in pgvector accepts up to 16000 dimensions, but both index types refuse to build above 2000. Attempting to create an HNSW or IVFFlat index on a wider column fails with an error stating the column cannot have more than 2000 dimensions.
Queries still work without an index; Postgres falls back to a sequential scan that computes the distance for every row, which is correct and unusably slow past a few tens of thousands of rows. Because it is only a performance cliff, this is often discovered in production rather than in development.
The options are to reduce the dimension at embedding time, which the Matryoshka models support directly, or to use the halfvec type, which stores half-precision values and supports indexing at higher dimensions with a modest accuracy cost. Decide before loading the corpus. Also remember that an IVFFlat index should be built after the table has representative data, since its clustering is derived from what is present at build time.