You're viewing version 2.17 of the OpenSearch documentation. This version is no longer maintained. For the latest version, see the current documentation. For information about OpenSearch version maintenance, see Release Schedule and Maintenance Policy.
Disk-based vector search
Introduced 2.17
For low-memory environments, OpenSearch provides disk-based vector search, which significantly reduces the operational costs for vector workloads. Disk-based vector search uses binary quantization, compressing vectors and thereby reducing the memory requirements. This memory optimization provides large memory savings at the cost of slightly increased search latency while still maintaining strong recall.
To use disk-based vector search, set the mode
parameter to on_disk
for your vector field type. This parameter will configure your index to use secondary storage.
Creating an index for disk-based vector search
To create an index for disk-based vector search, send the following request:
PUT my-vector-index
{
"settings" : {
"index": {
"knn": true
}
},
"mappings": {
"properties": {
"my_vector_field": {
"type": "knn_vector",
"dimension": 8,
"space_type": "innerproduct",
"data_type": "float",
"mode": "on_disk"
}
}
}
}
By default, the on_disk
mode configures the index to use the faiss
engine and hnsw
method. The default compression_level
of 32x
reduces the amount of memory the vectors require by a factor of 32. To preserve the search recall, rescoring is enabled by default. A search on a disk-optimized index runs in two phases: The compressed index is searched first, and then the results are rescored using full-precision vectors loaded from disk.
To reduce the compression level, provide the compression_level
parameter when creating the index mapping:
PUT my-vector-index
{
"settings" : {
"index": {
"knn": true
}
},
"mappings": {
"properties": {
"my_vector_field": {
"type": "knn_vector",
"dimension": 8,
"space_type": "innerproduct",
"data_type": "float",
"mode": "on_disk",
"compression_level": "16x"
}
}
}
}
For more information about the compression_level
parameter, see Compression levels. Note that for 4x
compression, the lucene
engine will be used.
If you need more granular fine-tuning, you can override additional k-NN parameters in the method definition. For example, to improve recall, increase the ef_construction
parameter value:
PUT my-vector-index
{
"settings" : {
"index": {
"knn": true
}
},
"mappings": {
"properties": {
"my_vector_field": {
"type": "knn_vector",
"dimension": 8,
"space_type": "innerproduct",
"data_type": "float",
"mode": "on_disk",
"method": {
"params": {
"ef_construction": 512
}
}
}
}
}
}
The on_disk
mode only works with the float
data type.
Ingestion
You can perform document ingestion for a disk-optimized vector index in the same way as for a regular vector index. To index several documents in bulk, send the following request:
POST _bulk
{ "index": { "_index": "my-vector-index", "_id": "1" } }
{ "my_vector_field": [1.5, 1.5, 1.5, 1.5, 1.5, 1.5, 1.5, 1.5], "price": 12.2 }
{ "index": { "_index": "my-vector-index", "_id": "2" } }
{ "my_vector_field": [2.5, 2.5, 2.5, 2.5, 2.5, 2.5, 2.5, 2.5], "price": 7.1 }
{ "index": { "_index": "my-vector-index", "_id": "3" } }
{ "my_vector_field": [3.5, 3.5, 3.5, 3.5, 3.5, 3.5, 3.5, 3.5], "price": 12.9 }
{ "index": { "_index": "my-vector-index", "_id": "4" } }
{ "my_vector_field": [4.5, 4.5, 4.5, 4.5, 4.5, 4.5, 4.5, 4.5], "price": 1.2 }
{ "index": { "_index": "my-vector-index", "_id": "5" } }
{ "my_vector_field": [5.5, 5.5, 5.5, 5.5, 5.5, 5.5, 5.5, 5.5], "price": 3.7 }
{ "index": { "_index": "my-vector-index", "_id": "6" } }
{ "my_vector_field": [6.5, 6.5, 6.5, 6.5, 6.5, 6.5, 6.5, 6.5], "price": 10.3 }
{ "index": { "_index": "my-vector-index", "_id": "7" } }
{ "my_vector_field": [7.5, 7.5, 7.5, 7.5, 7.5, 7.5, 7.5, 7.5], "price": 5.5 }
{ "index": { "_index": "my-vector-index", "_id": "8" } }
{ "my_vector_field": [8.5, 8.5, 8.5, 8.5, 8.5, 8.5, 8.5, 8.5], "price": 4.4 }
{ "index": { "_index": "my-vector-index", "_id": "9" } }
{ "my_vector_field": [9.5, 9.5, 9.5, 9.5, 9.5, 9.5, 9.5, 9.5], "price": 8.9 }
Search
Search is also performed in the same way as in other index configurations. The key difference is that, by default, the oversample_factor
of the rescore parameter is set to 3.0
(unless you override the compression_level
). For more information, see Rescoring quantized results using full precision. To perform vector search on a disk-optimized index, provide the search vector:
GET my-vector-index/_search
{
"query": {
"knn": {
"my_vector_field": {
"vector": [1.5, 2.5, 3.5, 4.5, 5.5, 6.5, 7.5, 8.5],
"k": 5
}
}
}
}
Similarly to other index configurations, you can override k-NN parameters in the search request:
GET my-vector-index/_search
{
"query": {
"knn": {
"my_vector_field": {
"vector": [1.5, 2.5, 3.5, 4.5, 5.5, 6.5, 7.5, 8.5],
"k": 5,
"method_parameters": {
"ef_search": 512
},
"rescore": {
"oversample_factor": 10.0
}
}
}
}
}
Radial search does not support disk-based vector search.
Model-based indexes
For model-based indexes, you can specify the on_disk
parameter in the training request in the same way that you would specify it during index creation. By default, on_disk
mode will use the Faiss IVF method and a compression level of 32x
. To run the training API, send the following request:
POST /_plugins/_knn/models/test-model/_train
{
"training_index": "train-index-name",
"training_field": "train-field-name",
"dimension": 8,
"max_training_vector_count": 1200,
"search_size": 100,
"description": "My model",
"space_type": "innerproduct",
"mode": "on_disk"
}
This command assumes that training data has been ingested into the train-index-name
index. For more information, see Building a k-NN index from a model.
You can override the compression_level
for disk-optimized indexes in the same way as for regular k-NN indexes.
Next steps
- For more information about binary quantization, see Binary quantization.
- For more information about k-NN vector workload modes, see Vector workload modes.