CREATE INDEX with diskann

Create a StreamingDiskANN index for high-performance vector search

Create a StreamingDiskANN index on a vector column for high-performance similarity search with optional label-based filtering.

• Create indexes for fast approximate nearest neighbor search
• Configure index build-time parameters for optimal performance
• Enable label-based filtering for precise vector search
• Choose storage layouts for memory optimization or plain storage

Note that:

• The index build process can be memory-intensive. Consider increasing maintenance_work_mem:

  SET maintenance_work_mem = '2GB';
  

• Label values must be within smallint range (-32768 to 32767)

• Creating indexes on UNLOGGED tables is not currently supported

• Null vectors are not indexed; null labels are treated as empty arrays

Samples

Basic index creation

CREATE INDEX document_embedding_idx ON document_embedding
USING diskann (embedding vector_cosine_ops);

With custom parameters

CREATE INDEX document_embedding_idx ON document_embedding
USING diskann (embedding vector_cosine_ops)
WITH (num_neighbors = 50, search_list_size = 100);

With label-based filtering

CREATE INDEX document_embedding_idx ON document_embedding
USING diskann (embedding vector_cosine_ops, labels);

With storage layout

CREATE INDEX document_embedding_idx ON document_embedding
USING diskann (embedding vector_cosine_ops)
WITH (storage_layout = 'memory_optimized');

Syntax

CREATE INDEX index_name ON table_name
USING diskann (embedding_column distance_ops [, labels_column])
[WITH (parameter = value, ...)];

Distance operators

Operator	Description	Use with
vector_cosine_ops	Cosine distance (<=>)	Normalized embeddings
vector_l2_ops	L2 distance (<->)	Euclidean distance
vector_ip_ops	Inner product (<#>)	Dot product (not compatible with plain storage)

Build-time parameters

Parameter	Type	Default	Description
storage_layout	text	memory_optimized	memory_optimized uses Statistical Binary Quantization; plain stores uncompressed data
num_neighbors	int	50	Maximum number of neighbors per node. Higher values increase accuracy but slow graph traversal
search_list_size	int	100	Search parameter during construction. Higher values improve graph quality but slow index builds
max_alpha	float	1.2	Alpha parameter in the algorithm. Higher values improve graph quality but slow index builds
num_dimensions	int	0 (all dimensions)	Number of dimensions to index. Useful for Matryoshka embeddings
num_bits_per_dimension	int	2 (if less than 900 dims), 1 otherwise	Bits per dimension when using Statistical Binary Quantization

For detailed information about each parameter and tuning recommendations, see Index build-time parameters.

To configure query behavior at runtime, see Query-time parameters.