Working with Vector Search

      +
      Use Vector Search with Full Text Search and Query.

      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.

              // Create a vector index configuration with a document property named "vector", 
              // 3 dimensions, and 100 centroids. Customize the encoding, the distance metric,
              // the number of probes, and the training size.
              var config = VectorIndexConfiguration(expression: "vector", dimensions: 3, centroids: 100)
              config.encoding = .none
              config.metric = .cosine
              config.numProbes = 8
              config.minTrainingSize = 2500
              config.maxTrainingSize = 5000

      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 to 3.

      • The amount of centroids is set to 100. 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, in this case colors_index, 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.

      Table 1. Vector Index Configuration Options
      Configuration Name Is Required Default Configuration Further Information

      Expression

      yes

      No default

      A SQL++ expression indicating where to get the vectors. A document property for embedded vectors or prediction() to call a registered Predictive model.

      Number of Dimensions

      yes

      No default

      2-4096

      Number of Centroids

      yes

      No default

      1-64000. The general guideline is an approximate square root of the number of documents

      Distance Metric

      no

      Squared Euclidean Distance (euclideanSquared)

      You can set the following alternates as your Distance Metric:

      • cosine (1 - Cosine similarity)

      • Euclidean

      • dot (negated dot product)

      Encoding

      no

      Scalar Quantizer(SQ) or SQ-8 bits

      There are three possible configurations:

      • None No compression, No data loss

      • Scalar Quantizer (SQ) or SQ-8 bits (Default) Reduces the number of bits per dimension

      • Product Quantizer (PQ) Reduces the number of dimensions and bits per dimension

      Training Size

      no

      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:

      • The minimum training size is set to 25x the number of Centroids or 2 PQ’s bits when PQ is used

      • The maximum training size is set to 256x the number of Centroids or 2 PQ’s bits when PQ is used

      NumProbes

      no

      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

      no

      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:

      1. You can call a Machine Learning(ML) model, and embed the generated vectors inside the documents.

      2. You can use the prediction() function to generate vectors to be indexed for each document at the indexing time.

      3. 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.

              // Get the collection named "colors" in the default scope.
              let collection = try database.collection(name: "colors")!;
              
              // Create a vector index configuration with a document property named "vector",
              // 3 dimensions, and 100 centroids.
              let config = VectorIndexConfiguration(expression: "vector", dimensions: 3, centroids: 100)
              // Create a vector index from the configuration with the name "colors_index".
              try collection.createIndex(withName: "colors_index", config: config)
      1. First, create the standard configuration, setting up an expression, number of dimensions and number of centroids for the vector embedding.

      2. 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.

          class ColorModel: PredictiveModel {
              func predict(input: DictionaryObject) -> DictionaryObject? {
                  // Get the color input from the input dictionary
                  guard let color = input.string(forKey: "colorInput") else {
                      fatalError("No input color found")
                  }
                  
                  // Use ML model to get a vector (an array of floats) for the input color.
                  guard let vector = try! Color.getVector(color: color) else {
                      return nil
                  }
                  
                  // Create an output dictionary by setting the vector result to
                  // the dictionary key named "vector".
                  let output = MutableDictionaryObject()
                  output.setValue(vector, forKey: "vector")
                  return output
              }
          }
          
          func createVectorIndexFromPredictiveIndex() throws {
              // Register the predictive model named "ColorModel".
              Database.prediction.registerModel(ColorModel(), withName: "ColorModel")
              
              // Create a vector index configuration with an expression using the prediction
              // function to get the vectors from the registered predictive model.
              let expression = "prediction(ColorModel, {\"colorInput\": color}).vector"
              let config = VectorIndexConfiguration(expression: expression, dimensions: 3, centroids: 100)
              
              // Create vector index from the configuration
              try collection.createIndex(withName: "colors_index", config: config)
          }
      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.

              // Creating a lazy vector index using the document's property named "color".
              // The "color" property's value will be used to compute a vector when updating the index.
              var config = VectorIndexConfiguration(expression: "color", dimensions: 3, centroids: 100)
              config.isLazy = true;

      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.

              guard let index = try collection.index(withName: "colors_index") else {
                  throw AppError.indexNotFound
              }
              
              while (true) {
                  // Start an update on it (in this case, limit to 50 entries at a time)
                  guard let updater = try index.beginUpdate(limit: 50) else {
                      // If updater is nil and no error, that means there are no more entries to process
                      break
                  }
                  
                  for i in 0..<updater.count {
                      // The value type will depend on the expression you have set in your index.
                      // In this example, it is a string property.
                      let color = updater.string(at: i)!
                      
                      var vector: [Float]? = nil
                      do {
                          vector = try await Color.getVectorAsync(color: color)
                      } catch ColorError.transient {
                          // Bad connection? Corrupted over the wire? Something bad happened
                          // and the vector cannot be generated at the moment. So skip
                          // this entry. The next time beginUpdate(limit:) is called,
                          // it will be considered again.
                          updater.skipVector(at: i)
                      }
                      
                      // Set the computed vector here. If vector is nil, calling setVector
                      // will cause the underlying document to NOT be indexed.
                      try updater.setVector(vector, at: i)
                  }
                  
                  // This writes the vectors to the index. You MUST have either set or
                  // skipped all the values inside the updater or this call will throw an error.
                  try updater.finish()
              }

      You procedurally update the vectors in the index by looping through the vectors in batches equivalent until you reach the value of the limit parameter.

      The update process follows the following sequence:

      1. Get a value for the updater.

        1. 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.
      2. Set the vector from the computed vector derived from the updater value and your ML model.

        1. If there is no value for the vector, this will result in the underlying document to not be indexed.

      3. 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.

      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 is an example of using Hybrid Search with the Full Text match() function.

              // Create a hybrid vector search query with full-text's match() that
              // uses the the full-text index named "color_desc_index".
              let sql = "SELECT meta().id, color " +
                        "WHERE MATCH(color_desc_index, $text) " +
                        "ORDER BY approx_vector_distance(vector, $vector) " +
                        "LIMIT 8"
              
              let query = try database.createQuery(sql)
              
              // Get a vector, an array of float numbers, for the input color code (e.g. FF000AA).
              // Normally, you will get the vector from your ML model.
              guard let vector = try Color.getVector(color: "FF00AA") else {
                  throw AppError.vectorNotFound
              }
              
              let parameters = Parameters()
              // Set the vector array to the parameter "$vector"
              parameters.setValue(vector, forName: "vector")
              // Set the vector array to the parameter "$text".
              parameters.setString("vibrant", forName: "text")
              query.parameters = parameters
              
              // Execute the query
              let results = try query.execute()
              
              for r in results {
                  // Process result
              }

      Below is an example of using Hybrid Search with an array of vectors generated by the Prediction() function at index time.

              // Create a hybrid vector search query that uses prediction() for computing vectors.
              let sql = 
              "SELECT meta().id, color " +
              "WHERE saturation > 0.5 " +
              "ORDER BY approx_vector_distance(prediction(ColorModel, {\"colorInput\": color}).vector, $vector) " +
              "LIMIT 8"
              
              let query = try database.createQuery(sql)
              
              // Get a vector, an array of float numbers, for the input color code (e.g. FF000AA).
              // Normally, you will get the vector from your ML model.
              guard let vector = try Color.getVector(color: "FF00AA") else {
                  throw AppError.vectorNotFound
              }
              
              // Set the vector array to the parameter "$vector"
              let parameters = Parameters()
              parameters.setValue(vector, forName: "vector")
              query.parameters = parameters
              
              // Execute the query
              let results = try query.execute()
              
              for r in results {
                  // Process result
              }

      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

      yes

      The expression returning a vector (NOT Index Name). Must match the expression specified in the vector index exactly.

      target-vector

      yes

      The target vector.

      metric

      no

      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

      no

      Number of buckets to search for the nearby vectors. If not specified, the nprobes set in the vector index is used.

      accurate

      no

      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()

              // Create a vector search query by using the approx_vector_distance() in WHERE clause.
              let sql = "SELECT meta().id, color " +
                        "FROM _default.colors " +
                        "WHERE approx_vector_distance(vector, $vector) < 0.5 " +
                        "LIMIT 8"
              
              let query = try database.createQuery(sql)
              
              // Get a vector, an array of float numbers, for the input color code (e.g. FF000AA).
              // Normally, you will get the vector from your ML model.
              guard let vector = try Color.getVector(color: "FF00AA") else {
                  throw AppError.vectorNotFound
              }
              
              // Set the vector array to the parameter "$vector"
              let parameters = Parameters()
              parameters.setValue(vector, forName: "vector")
              query.parameters = parameters
              
              // Execute the query
              let results = try query.execute()
              
              for r in results {
                  // Process result
              }

      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 is an example of using APPROX_VECTOR_DISTANCE() with an array of vectors generated by the Prediction() function at index time.

              // Create a vector search query that uses prediction() for computing vectors.
              let sql =
              "SELECT id, color " +
              "FROM _default.colors " +
              "ORDER BY approx_vector_distance(prediction(ColorModel, {\"colorInput\": color}).vector, $vector) " +
              "LIMIT 8"
              
              let query = try database.createQuery(sql)
              
              // Get a vector, an array of float numbers, for the input color code (e.g. FF000AA).
              // Normally, you will get the vector from your ML model.
              guard let vector = try Color.getVector(color: "FF00AA") else {
                  throw AppError.vectorNotFound
              }
              
              // Set the vector array to the parameter "$vector"
              let parameters = Parameters()
              parameters.setValue(vector, forName: "vector")
              query.parameters = parameters
              
              // Execute the query
              let results = try query.execute()
              
              for r in results {
                  // Process result
              }