Use Vector Search with Full Text Search and Query.
Use Vector Search
To configure a project to use vector search, follow the installation instructions to add the Vector Search extension.
You must install Couchbase Lite to use the Vector Search extension. |
Create a Vector Index
This method shows how you can create a vector index using the Couchbase Lite Vector Search extension.
-
Java
-
Kotlin
// create the configuration for a vector index named "vector"
// with 3 dimensions, 100 centroids, no encoding, using cosine distance
// with a max training size 5000 and amin training size 2500
// no vector encoding and using COSINE distance measurement
VectorIndexConfiguration config = new VectorIndexConfiguration("vector", 3L, 100L)
.setEncoding(VectorEncoding.none())
.setMetric(VectorIndexConfiguration.DistanceMetric.COSINE)
.setNumProbes(8L)
.setMinTrainingSize(2500L)
.setMaxTrainingSize(5000L);
First, initialize the config
object with the VectorIndexConfiguration()
method with the following parameters:
-
The expression of the data as a vector.
-
The width or
dimensions
of the vector index is set to3
. -
The amount of
centroids
is set to100
. This means that there will be one hundred buckets with a single centroid each that gathers together similar vectors.
You can also alter some optional config settings such as encoding
.
From there, you create an index within a given collection using the previously generated config
object.
// create the configuration for a vector index named "vector"
// with 3 dimensions, 100 centroids, no encoding, using cosine distance
// with a max training size 5000 and amin training size 2500
// no vector encoding and using COSINE distance measurement
val config = VectorIndexConfigurationFactory.newConfig(
"vector",
3L,
100L,
encoding = VectorEncoding.none(),
metric = VectorIndexConfiguration.DistanceMetric.COSINE,
numProbes = 8L,
minTrainingSize = 2500L,
maxTrainingSize = 5000L
)
First, initialize the config
object with the VectorIndexConfiguration()
method with the following parameters:
-
The expression of the data as a vector.
-
The width or
dimensions
of the vector index is set to3
. -
The amount of
centroids
is set to100
. This means that there will be one hundred buckets with a single centroid each that gathers together similar vectors.
You can also alter some optional config settings such as encoding
.
From there, you create an index within a given collection using the previously generated config
object.
The number of vectors, the width or dimensions of the vectors and the training size can incur high CPU and memory costs as the size of each variable increases. This is because the training vectors have to be resident on the machine. |
Vector Index Configuration
The table below displays the different configurations you can modify within your VectorIndexConfiguration()
function.
For more information on specific configurations, see Vector Search.
Configuration Name | Is Required | Default Configuration | Further Information |
---|---|---|---|
Expression |
No default |
A SQL++ expression indicating where to get the vectors.
A document property for embedded vectors or
|
|
Number of Dimensions |
No default |
2-4096 |
|
Number of Centroids |
No default |
1-64000. The general guideline is an approximate square root of the number of documents |
|
Distance Metric |
Squared Euclidean Distance (euclideanSquared) |
You can set the following alternates as your Distance Metric:
|
|
Encoding |
Scalar Quantizer(SQ) or SQ-8 bits |
There are three possible configurations:
|
|
Training Size |
The default values for both the minimum and maximum training size is zero. The training size is calculated based on the number of Centroids and the encoding type. |
The guidelines for the minimum and maximum training size are as follows:
|
|
NumProbes |
The default value is 0. The number of Probes is calculated based on the number of Centroids |
A guideline for setting a custom number of probes is at least 8 or 0.5% the number of Centroids |
|
isLazy |
False |
Setting the value to true will enable lazy mode for the vector index |
Altering the default training sizes could be detrimental to the accuracy of returned results produced by the model and total computation time. |
Generating Vectors
You can use the following methods to generate vectors in Couchbase Lite:
-
You can call a Machine Learning(ML) model, and embed the generated vectors inside the documents.
-
You can use the
prediction()
function to generate vectors to be indexed for each document at the indexing time. -
You can use Lazy Vector Index (lazy index) to generate vectors asynchronously from remote ML models that may not always be reachable or functioning, skipping or scheduling retries for those specific cases.
Below are example configurations of the previously mentioned methods.
Create a Vector Index with Embeddings
This method shows you how to create a Vector Index with embeddings.
-
Java
-
Kotlin
// create a vector index named "colors_index"
// in the collection "_default.colors"
db.getCollection("colors").createIndex(
"colors_index",
new VectorIndexConfiguration("vector", 3L, 100L));
-
First, create the standard configuration, setting up an expression, number of dimensions and number of centroids for the vector embedding.
-
Next, create a vector index,
colors_index
, on a collection and pass it the configuration.
// create a vector index named "colors_index"
// in the collection "_default.colors"
db.getCollection("colors")?.createIndex(
"colors_index",
VectorIndexConfigurationFactory.newConfig("vector", 3L, 100L)
) ?: throw IllegalStateException("No such collection: colors")
-
First, create the standard configuration, setting up an expression, number of dimensions and number of centroids for the vector embedding.
-
Next, create a vector index,
colors_index
, on a collection and pass it the configuration.
Create Vector Index Embeddings from a Predictive Model
This method generates vectors to be indexed for each document at the index time by using the prediction()
function.
The key difference to note is that the config
object uses the output of the prediction()
function as the expression
parameter to generate the vector index.
-
Java
-
Kotlin
// create a vector index with a simple predictive model
Database.prediction.registerModel("ColorModel", colorModel);
db.getCollection("colors").createIndex(
"colors_pred_index",
new VectorIndexConfiguration(
"prediction(ColorModel, {'colorInput': color}).vector",
3L, 100L));
// create a vector index with a simple predictive model
Database.prediction.registerModel("ColorModel", colorModel)
db.getCollection("colors")?.createIndex(
"colors_pred_index",
VectorIndexConfigurationFactory.newConfig(
"prediction(ColorModel, {'colorInput': color}).vector",
3L, 100L
)
) ?: throw IllegalStateException("No such collection: colors")
You can use less storage by using the prediction() function as the encoded vectors will only be stored in the index.
However, the index time will be longer as vector embedding generation is occurring at run time.
|
Create a Lazy Vector Index
Lazy indexing is an alternate approach to using the standard predictive model with regular vector indexes which handle the indexing process automatically. You can use lazy indexing to use a ML model that is not available locally on the device and to create vector indexes without having vector embeddings in the documents.
-
Java
-
Kotlin
db.getCollection("colors").createIndex(
"colors_index",
new VectorIndexConfiguration("color", 3L, 100L)
.setLazy(true));
You can enable lazy vector indexing by setting the isLazy
property to true
in your vector index configuration.
db.getCollection("colors")?.createIndex(
"colors_index",
VectorIndexConfigurationFactory.newConfig("color", 3L, 100L, lazy = true)
) ?: throw IllegalStateException("No such collection: colors")
You can enable lazy vector indexing by setting the isLazy
property to true
in your vector index configuration.
Lazy Vector Indexing is opt-in functionality, the isLazy property is set to false by default.
|
Updating the Lazy Index
Below is an example of how you can update your lazy index.
-
Java
-
Kotlin
while (true) {
try (IndexUpdater updater = col.getIndex("colors_index").beginUpdate(10)) {
if (updater == null) { break; }
for (int i = 0; i < updater.count(); i++) {
try {
// get the color swatch from the updater and send it to the remote model
List<Float> embedding = colorModel.getEmbedding(updater.getBlob(i));
updater.setVector(embedding, i);
}
catch (IOException e) {
// Bad connection? Corrupted over the wire? Something bad happened
// and the vector cannot be generated at the moment: skip it.
// The next time beginUpdate() is called, we'll try it again.
updater.skipVector(i);
}
}
// This writes the vectors to the index. You MUST either have set or skipped each
// of the the vectors in the updater or this call will throw an exception.
updater.finish();
}
}
You procedurally update the vectors in the index by looping through the vectors in batches until you reach the value of the limit
parameter.
The update process follows the following sequence:
-
Get a value for the updater.
-
If the there is no value for the vector, handle it. In this case, the vector will be skipped and considered the next time
beginUpdate()
is called.A key benefit of lazy indexing is that the indexing process continues if a vector fails to generate. For standard vector indexing, this will cause the affected documents to be dropped from the indexing process.
-
-
Set the vector from the computed vector derived from the updater value and your ML model.
-
If there is no value for the vector, this will result in the underlying document to not be indexed.
-
-
Once all vectors have completed the update loop, finish updating.
while (true) {
col.getIndex("colors_index")?.beginUpdate(10)?.use { updater ->
for (i in 0 until updater.count()) {
try {
val embedding: List<Float?>? = colorModel.getEmbedding(updater.getBlob(i))
updater.setVector(embedding, i)
} catch (e: IOException) {
// Bad connection? Corrupted over the wire? Something bad happened
// and the vector cannot be generated at the moment: skip it.
// The next time beginUpdate() is called, we'll try it again.
updater.skipVector(i)
}
}
// This writes the vectors to the index. You MUST either have set or skipped each
// of the the vectors in the updater or this call will throw an exception.
updater.finish()
}
// loop until there are no more vectors to update
?: break
}
You procedurally update the vectors in the index by looping through the vectors in batches until you reach the value of the limit
parameter.
The update process follows the following sequence:
-
Get a value for the updater.
-
If the there is no value for the vector, handle it. In this case, the vector will be skipped and considered the next time
beginUpdate()
is called.A key benefit of lazy indexing is that the indexing process continues if a vector fails to generate. For standard vector indexing, this will cause the affected documents to be dropped from the indexing process.
-
-
Set the vector from the computed vector derived from the updater value and your ML model.
-
If there is no value for the vector, this will result in the underlying document to not be indexed.
-
-
Once all vectors have completed the update loop, finish updating.
updater.finish() will throw an error if any values inside the updater have not been set or skipped.
|
Vector Search SQL++ Support
Couchbase Lite currently supports Hybrid Vector Search and the APPROX_VECTOR_DISTANCE()
function.
Similar to the Full Text Search match() function, the APPROX_VECTOR_DISTANCE() function and Hybrid Vector Search cannot use the OR expression with the other expressions in the related WHERE clause.
|
Use Hybrid Vector Search
You can use Hybrid Vector Search (Hybrid Search) to perform vector search in conjunction with regular SQL++ queries.
With Hybrid Search, you perform vector search on documents that have already been filtered based on criteria specified in the WHERE
clause.
A LIMIT clause is required for non-hybrid Vector Search, this avoids a slow, exhaustive unlimited search of all possible vectors.
|
Hybrid Vector Search with Full Text Match
Below are examples of using Hybrid Search with the Full Text match()
function.
-
Java
-
Kotlin
// Create a hybrid vector search query with full-text's match() that
// uses the the full-text index named "color_desc_index".
Query query = db.createQuery(
"SELECT meta().id, color"
+ " FROM _default.colors"
+ " WHERE MATCH(color_desc_index, $text)"
+ " ORDER BY APPROX_VECTOR_DISTANCE(vector, $vector)"
+ " LIMIT 8");
Parameters params = new Parameters();
params.setArray("vectorParam", new MutableArray(colorVector));
query.setParameters(params);
try (ResultSet rs = query.execute()) {
// process results
}
// Create a hybrid vector search query with full-text's match() that
// uses the the full-text index named "color_desc_index".
val query = db.createQuery(
("SELECT meta().id, color"
+ " FROM _default.colors"
+ " WHERE MATCH(color_desc_index, \$text)"
+ " ORDER BY APPROX_VECTOR_DISTANCE(vector, \$vector)"
+ " LIMIT 8")
)
val params = Parameters()
params.setArray("vectorParam", MutableArray((colorVector)))
query.parameters = params
query.execute().use { rs ->
// process results
}
Prediction with Hybrid Vector Search
Below are examples of using Hybrid Search with an array of vectors generated by the Prediction()
function at index time.
-
Java
-
Kotlin
Query query = db.createQuery(
"SELECT meta().id, color"
+ " FROM _default.colors"
+ " WHERE saturation > 0.5"
+ " ORDER BY APPROX_VECTOR_DISTANCE("
+ " prediction(ColorModel, {'colorInput': color}).vector,"
+ " $vectorParam)"
+ " LIMIT 8");
Parameters params = new Parameters();
params.setArray("vectorParam", new MutableArray(colorVector));
query.setParameters(params);
try (ResultSet rs = query.execute()) {
// process results
}
val query = db.createQuery(
("SELECT meta().id, color"
+ " FROM _default.colors"
+ " WHERE saturation > 0.5"
+ " ORDER BY APPROX_VECTOR_DISTANCE("
+ " prediction(ColorModel, {'colorInput': color}).vector,"
+ " \$vectorParam)"
+ " LIMIT 8")
)
val params = Parameters()
params.setArray("vectorParam", MutableArray((colorVector)))
query.parameters = params
query.execute().use { rs ->
// process results
}
APPROX_VECTOR_DISTANCE(vector-expr, target-vector, [metric], [nprobes], [accurate])
If you use a different distance metric in the APPROX_VECTOR_DISTANCE() function from the one configured in the index, you will receive an error when compiling the query.
|
Parameter | Is Required | Description |
---|---|---|
vector-expr |
The expression returning a vector (NOT Index Name). Must match the expression specified in the vector index exactly. |
|
target-vector |
The target vector. |
|
metric |
Values : "EUCLIDEAN_SQUARED", “L2_SQUARED”, “EUCLIDEAN”, “L2”, ”COSINE”, “DOT”. If not specified, the metric set in the vector index is used. If specified, the metric must match with the metric set in the vector index. This optional parameter allows multiple indexes to be attached to the same field in a document. |
|
nprobes |
Number of buckets to search for the nearby vectors. If not specified, the nprobes set in the vector index is used. |
|
accurate |
If not present, false will be used, which means that the quantized/encoded vectors in the index will be used for calculating the distance. IMPORTANT: Only accurate = false is supported |
Use APPROX_VECTOR_DISTANCE()
-
Java
-
Kotlin
// use APPROX_VECTOR_DISTANCE in a query WHERE clause
Query query = db.createQuery(
"SELECT meta().id, color"
+ " FROM _default.colors"
+ " WHERE APPROX_VECTOR_DISTANCE(vector, $vectorParam) < 0.5");
Parameters params = new Parameters();
params.setArray("vectorParam", new MutableArray(colorVector));
query.setParameters(params);
try (ResultSet rs = query.execute()) {
// process results
}
This function returns the approximate distance between a given vector, typically generated from your ML model, and an array of vectors with size equal to the LIMIT
parameter, collected by a SQL++ query using APPROX_VECTOR_DISTANCE()
.
// use APPROX_VECTOR_DISTANCE in a query WHERE clause
val query = db.createQuery(
("SELECT meta().id, color"
+ " FROM _default.colors"
+ " WHERE APPROX_VECTOR_DISTANCE(vector, \$vectorParam) < 0.5")
)
val params = Parameters()
params.setArray("vectorParam", MutableArray((colorVector)))
query.parameters = params
query.execute().use { rs ->
// process results
}
This function returns the approximate distance between a given vector, typically generated from your ML model, and an array of vectors with size equal to the LIMIT
parameter, collected by a SQL++ query using APPROX_VECTOR_DISTANCE()
.
Prediction with APPROX_VECTOR_DISTANCE()
Below are examples of using APPROX_VECTOR_DISTANCE()
with an array of vectors generated by the Prediction()
function at index time.
-
Java
-
Kotlin
// use APPROX_VECTOR_DISTANCE with a predictive model
Database.prediction.registerModel("ColorModel", colorModel);
db.getCollection("colors").createIndex(
"colors_pred_index",
new VectorIndexConfiguration(
"prediction(ColorModel, {'colorInput': color}).vector",
3L, 100L));
Query query = db.createQuery(
"SELECT meta().id, color"
+ " FROM _default.colors"
+ " ORDER BY APPROX_VECTOR_DISTANCE("
+ " prediction(ColorModel, {'colorInput': color}).vector,"
+ " $vectorParam)"
+ " LIMIT 300");
Parameters params = new Parameters();
params.setArray("vectorParam", new MutableArray(colorVector));
query.setParameters(params);
try (ResultSet rs = query.execute()) {
// process results
}
// use APPROX_VECTOR_DISTANCE with a predictive model
Database.prediction.registerModel("ColorModel", (colorModel))
db.getCollection("colors")?.createIndex(
"colors_pred_index",
VectorIndexConfigurationFactory.newConfig(
"prediction(ColorModel, {'colorInput': color}).vector",
3L, 100L
)
) ?: throw IllegalStateException("No such collection: colors")
val query = db.createQuery(
("SELECT meta().id, color"
+ " FROM _default.colors"
+ " ORDER BY APPROX_VECTOR_DISTANCE("
+ " prediction(ColorModel, {'colorInput': color}).vector,"
+ " \$vectorParam)"
+ " LIMIT 300")
)
val params = Parameters()
params.setArray("vectorParam", MutableArray((colorVector)))
query.parameters = params
query.execute().use { rs ->
// process results
}