Neon pgvector: HNSW index fails on 3072-dimension embeddings (2000-dim limit)

[field troubleshooting]: do not ANN-index vector(3072) directly: use text-embedding-3-small (1536 dims) or another <=2000-dim representation, or switch the column to halfvec to halve the dimension footprint. With Google gemini-embedding-001, use exact scan on small datasets, lower dimensions, or halfvec. Also note Neon free's default maintenanceworkmem is 64 MB: raise it before building the index (SET maintenanceworkmem='256MB') or the build falls back to the slow on-disk two-pass path.

Context: Creating an HNSW index on a vector(3072) column fails. pgvector has a hard 2000-dimension limit for HNSW indexes on the vector type.

Source: Creating an HNSW index on a vector(3072) column fails. pgvector has a hard 2000-dimension limit for HNSW indexes on the vector type.

Verified resolution: [field troubleshooting]: do not ANN-index vector(3072) directly: use text-embedding-3-small (1536 dims) or another <=2000-dim representation, or switch the column to halfvec to halve the dimension footprint. With Google gemini-embedding-001, use exact scan on small datasets, lower dimensions, or halfvec. Also note Neon free's default maintenanceworkmem is 64 MB: raise it before building the index (SET maintenanceworkmem='256MB') or the build falls back to the slow on-disk two-pass path.

Source URL: https://github.com/aurahq-ai/aura/blob/HEAD/content/blog/why-neon-postgres.mdx