Implement k-NN approximate search and mappings bindings (#215)

* Implement KNN bindings

Signed-off-by: Thomas Farr <tsfarr@amazon.com>

* Minimal docs

Signed-off-by: Thomas Farr <tsfarr@amazon.com>

* Add to changelog

Signed-off-by: Thomas Farr <tsfarr@amazon.com>

---------

Signed-off-by: Thomas Farr <tsfarr@amazon.com>
(cherry picked from commit 3833005)

CHANGELOG.md

@@ -2,6 +2,10 @@
 Inspired from [Keep a Changelog](https://keepachangelog.com/en/1.0.0/)
 
 ## [Unreleased]
+
+### Added
+- Added support for approximate k-NN search queries and k-NN vector index properties ([#215](https://github.com/opensearch-project/opensearch-net/pull/215))
+
 ### Dependencies
 - Bumps `System.Reflection.Emit` from 4.3.0 to 4.7.0
 - Bumps `Argu` from 5.5.0 to 6.1.1

abstractions/src/OpenSearch.OpenSearch.Managed/OpenSearchNode.cs

@@ -29,6 +29,8 @@
 using System;
 using System.Collections.Generic;
 using System.Diagnostics;
+using System.IO;
+using System.Runtime.InteropServices;
 using System.Threading;
 using OpenSearch.OpenSearch.Managed.Configuration;
 using OpenSearch.OpenSearch.Managed.ConsoleWriters;
@@ -93,9 +95,25 @@ private static Dictionary<string, string> EnvVars(NodeConfiguration config)
 			if (!string.IsNullOrWhiteSpace(config.FileSystem.OpenSearchHome))
 				environmentVariables.Add("OPENSEARCH_HOME", config.FileSystem.OpenSearchHome);
 
+			var knnLibDir = Path.Combine(config.FileSystem.OpenSearchHome, "plugins", "opensearch-knn", config.Version.Major >= 2 ? "lib" : "knnlib");
+			if (RuntimeInformation.IsOSPlatform(OSPlatform.OSX))
+				AppendPathEnvVar("JAVA_LIBRARY_PATH", knnLibDir);
+			else if (RuntimeInformation.IsOSPlatform(OSPlatform.Linux))
+				AppendPathEnvVar("LD_LIBRARY_PATH", knnLibDir);
+
 			return environmentVariables;
 		}
 
+		private static void AppendPathEnvVar(string name, string value)
+		{
+			var previous = Environment.GetEnvironmentVariable(name);
+			Environment.SetEnvironmentVariable(name,
+				string.IsNullOrWhiteSpace(previous)
+					? value
+					: $"{previous}{Path.PathSeparator}{value}"
+				);
+		}
+
 		private bool AssumedStartedStateChecker(string section, string message)
 		{
 			if (AssumeStartedOnNotEnoughMasterPing

src/OpenSearch.Client/Mapping/DynamicTemplate/SingleMapping.cs

@@ -149,6 +149,10 @@ public IProperty Generic(Func<GenericPropertyDescriptor<T>, IGenericProperty> se
 		public IProperty SearchAsYouType(Func<SearchAsYouTypePropertyDescriptor<T>, ISearchAsYouTypeProperty> selector) =>
 			selector?.Invoke(new SearchAsYouTypePropertyDescriptor<T>());
 
+		/// <inheritdoc />
+		public IProperty KnnVector(Func<KnnVectorPropertyDescriptor<T>, IKnnVectorProperty> selector) =>
+			selector?.Invoke(new KnnVectorPropertyDescriptor<T>());
+
 #pragma warning disable CS3001 // Argument type is not CLS-compliant
 		public IProperty Scalar(Expression<Func<T, int>> field, Func<NumberPropertyDescriptor<T>, INumberProperty> selector = null) =>
 			selector.InvokeOrDefault(new NumberPropertyDescriptor<T>().Name(field).Type(NumberType.Integer));

src/OpenSearch.Client/Mapping/Types/FieldType.cs

@@ -159,6 +159,9 @@ public enum FieldType
 		RankFeature,
 
 		[EnumMember(Value = "rank_features")]
-		RankFeatures
+		RankFeatures,
+
+		[EnumMember(Value = "knn_vector")]
+		KnnVector
 	}
 }

src/OpenSearch.Client/Mapping/Types/Properties.cs

@@ -157,6 +157,9 @@ TReturnType Nested<TChild>(Func<NestedPropertyDescriptor<T, TChild>, INestedProp
 
 		/// <inheritdoc cref="ISearchAsYouTypeProperty"/>
 		TReturnType SearchAsYouType(Func<SearchAsYouTypePropertyDescriptor<T>, ISearchAsYouTypeProperty> selector);
+
+		/// <inheritdoc cref="IKnnVectorProperty" />
+		TReturnType KnnVector(Func<KnnVectorPropertyDescriptor<T>, IKnnVectorProperty> selector);
 	}
 
 	public partial class PropertiesDescriptor<T> where T : class
@@ -252,6 +255,9 @@ public PropertiesDescriptor<T> Object<TChild>(Func<ObjectTypeDescriptor<T, TChil
 		/// <inheritdoc cref="IRankFeaturesProperty"/>
 		public PropertiesDescriptor<T> RankFeatures(Func<RankFeaturesPropertyDescriptor<T>, IRankFeaturesProperty> selector) => SetProperty(selector);
 
+		/// <inheritdoc cref="IKnnVectorProperty" />
+		public PropertiesDescriptor<T> KnnVector(Func<KnnVectorPropertyDescriptor<T>, IKnnVectorProperty> selector) => SetProperty(selector);
+
 		/// <summary>
 		/// Map a custom property.
 		/// </summary>

src/OpenSearch.Client/Mapping/Types/PropertyFormatter.cs

@@ -118,6 +118,7 @@ public IProperty Deserialize(ref JsonReader reader, IJsonFormatterResolver forma
 				case FieldType.Alias: return Deserialize<FieldAliasProperty>(ref segmentReader, formatterResolver);
 				case FieldType.RankFeature: return Deserialize<RankFeatureProperty>(ref segmentReader, formatterResolver);
 				case FieldType.RankFeatures: return Deserialize<RankFeaturesProperty>(ref segmentReader, formatterResolver);
+				case FieldType.KnnVector: return Deserialize<KnnVectorProperty>(ref segmentReader, formatterResolver);
 				case FieldType.None:
 					// no "type" field in the property mapping, or FieldType enum could not be parsed from typeString
 					return Deserialize<ObjectProperty>(ref segmentReader, formatterResolver);

src/OpenSearch.Client/Mapping/Types/Specialized/Knn/KnnVectorProperty.cs

@@ -0,0 +1,163 @@
+/* SPDX-License-Identifier: Apache-2.0
+*
+* The OpenSearch Contributors require contributions made to
+* this file be licensed under the Apache-2.0 license or a
+* compatible open source license.
+*/
+
+using System;
+using System.Collections.Generic;
+using System.Diagnostics;
+using System.Runtime.Serialization;
+using OpenSearch.Net.Utf8Json;
+
+namespace OpenSearch.Client;
+
+[ReadAs(typeof(KnnVectorProperty))]
+[InterfaceDataContract]
+public interface IKnnVectorProperty : IDocValuesProperty
+{
+	/// <summary>
+	/// The dimension of the vector.
+	/// </summary>
+	[DataMember(Name = "dimension")]
+	int? Dimension { get; set; }
+
+	/// <summary>
+	/// The model to use when the underlying Approximate k-NN algorithm requires a training step.
+	/// </summary>
+	[DataMember(Name = "model_id")]
+	string ModelId { get; set; }
+
+	/// <summary>
+	/// The method to use when the underlying Approximate k-NN algorithm does not require training.
+	/// </summary>
+	[DataMember(Name = "method")]
+	IKnnMethod Method { get; set; }
+}
+
+[ReadAs(typeof(KnnMethod))]
+[InterfaceDataContract]
+public interface IKnnMethod
+{
+	/// <summary>
+	/// The identifier for the nearest neighbor method.
+	/// </summary>
+	[DataMember(Name = "name")]
+	string Name { get; set; }
+
+	/// <summary>
+	/// The approximate k-NN library to use for indexing and search.
+	/// </summary>
+	[DataMember(Name = "engine")]
+	string Engine { get; set; }
+
+	/// <summary>
+	/// The vector space used to calculate the distance between vectors.
+	/// </summary>
+	[DataMember(Name = "space_type")]
+	string SpaceType { get; set; }
+
+	/// <summary>
+	/// The parameters used for the nearest neighbor method.
+	/// </summary>
+	[DataMember(Name = "parameters")]
+	IDictionary<string, object> Parameters { get; set; }
+}
+
+public class KnnMethod : IKnnMethod
+{
+	/// <inheritdoc />
+	public string Name { get; set; }
+	/// <inheritdoc />
+	public string Engine { get; set; }
+	/// <inheritdoc />
+	public string SpaceType { get; set; }
+	/// <inheritdoc />
+	public IDictionary<string, object> Parameters { get; set; }
+}
+
+[InterfaceDataContract]
+[JsonFormatter(typeof(VerbatimDictionaryKeysFormatter<KnnMethodParameters, IKnnMethodParameters, string, object>))]
+public interface IKnnMethodParameters : IIsADictionary<string, object> { }
+
+public class KnnMethodParameters : IsADictionaryBase<string, object>, IKnnMethodParameters
+{
+	public KnnMethodParameters() { }
+
+	public KnnMethodParameters(IDictionary<string, object> container) : base(container) { }
+
+	public KnnMethodParameters(Dictionary<string, object> container) : base(container) { }
+
+	public void Add(string name, object value) => BackingDictionary.Add(name, value);
+}
+
+[DebuggerDisplay("{DebugDisplay}")]
+public class KnnVectorProperty : DocValuesPropertyBase, IKnnVectorProperty
+{
+	public KnnVectorProperty() : base(FieldType.KnnVector) { }
+
+	/// <inheritdoc />
+	public int? Dimension { get; set; }
+	/// <inheritdoc />
+	public string ModelId { get; set; }
+	/// <inheritdoc />
+	public IKnnMethod Method { get; set; }
+}
+
+[DebuggerDisplay("{DebugDisplay}")]
+public class KnnVectorPropertyDescriptor<T>
+	: DocValuesPropertyDescriptorBase<KnnVectorPropertyDescriptor<T>, IKnnVectorProperty, T>, IKnnVectorProperty
+	where T : class
+{
+	public KnnVectorPropertyDescriptor() : base(FieldType.KnnVector) { }
+
+	int? IKnnVectorProperty.Dimension { get; set; }
+	string IKnnVectorProperty.ModelId { get; set; }
+	IKnnMethod IKnnVectorProperty.Method { get; set; }
+
+	/// <inheritdoc cref="IKnnVectorProperty.Dimension" />
+	public KnnVectorPropertyDescriptor<T> Dimension(int? dimension) =>
+		Assign(dimension, (p, v) => p.Dimension = v);
+
+	/// <inheritdoc cref="IKnnVectorProperty.ModelId" />
+	public KnnVectorPropertyDescriptor<T> ModelId(string modelId) =>
+		Assign(modelId, (p, v) => p.ModelId = v);
+
+	/// <inheritdoc cref="IKnnVectorProperty.Method" />
+	public KnnVectorPropertyDescriptor<T> Method(Func<KnnMethodDescriptor, IKnnMethod> selector) =>
+		Assign(selector, (p, v) => p.Method = v?.Invoke(new KnnMethodDescriptor()));
+}
+
+public class KnnMethodDescriptor
+	: DescriptorBase<KnnMethodDescriptor, IKnnMethod>, IKnnMethod
+{
+	string IKnnMethod.Name { get; set; }
+	string IKnnMethod.Engine { get; set; }
+	string IKnnMethod.SpaceType { get; set; }
+	IDictionary<string, object> IKnnMethod.Parameters { get; set; }
+
+	/// <inheritdoc cref="IKnnMethod.Name" />
+	public KnnMethodDescriptor Name(string name) =>
+		Assign(name, (c, v) => c.Name = v);
+
+	/// <inheritdoc cref="IKnnMethod.Engine" />
+	public KnnMethodDescriptor Engine(string engine) =>
+		Assign(engine, (c, v) => c.Engine = v);
+
+	/// <inheritdoc cref="IKnnMethod.SpaceType" />
+	public KnnMethodDescriptor SpaceType(string spaceType) =>
+		Assign(spaceType, (c, v) => c.SpaceType = v);
+
+	/// <inheritdoc cref="IKnnMethod.Parameters" />
+	public KnnMethodDescriptor Parameters(Func<KnnMethodParametersDescriptor, IPromise<IKnnMethodParameters>> selector) =>
+		Assign(selector, (c, v) => c.Parameters = v?.Invoke(new KnnMethodParametersDescriptor())?.Value);
+}
+
+public class KnnMethodParametersDescriptor : IsADictionaryDescriptorBase<KnnMethodParametersDescriptor, IKnnMethodParameters, string, object>
+{
+	public KnnMethodParametersDescriptor() : base(new KnnMethodParameters()) { }
+
+	public KnnMethodParametersDescriptor Parameter(string name, object value) =>
+		Assign(name, value);
+}

src/OpenSearch.Client/Mapping/Visitor/IPropertyVisitor.cs

@@ -88,6 +88,8 @@ public interface IPropertyVisitor
 
 		void Visit(IFieldAliasProperty type, PropertyInfo propertyInfo, OpenSearchPropertyAttributeBase attribute);
 
+		void Visit(IKnnVectorProperty type, PropertyInfo propertyInfo, OpenSearchPropertyAttributeBase attribute);
+
 		IProperty Visit(PropertyInfo propertyInfo, OpenSearchPropertyAttributeBase attribute);
 
 		bool SkipProperty(PropertyInfo propertyInfo, OpenSearchPropertyAttributeBase attribute);

src/OpenSearch.Client/Mapping/Visitor/NoopPropertyVisitor.cs

@@ -89,6 +89,8 @@ public virtual void Visit(ISearchAsYouTypeProperty type, PropertyInfo propertyIn
 
 		public virtual void Visit(IFieldAliasProperty type, PropertyInfo propertyInfo, OpenSearchPropertyAttributeBase attribute) { }
 
+		public virtual void Visit(IKnnVectorProperty type, PropertyInfo propertyInfo, OpenSearchPropertyAttributeBase attribute) { }
+
 		public virtual IProperty Visit(PropertyInfo propertyInfo, OpenSearchPropertyAttributeBase attribute) => null;
 
 		public void Visit(IProperty type, PropertyInfo propertyInfo, OpenSearchPropertyAttributeBase attribute)
@@ -176,6 +178,9 @@ public void Visit(IProperty type, PropertyInfo propertyInfo, OpenSearchPropertyA
 				case IFieldAliasProperty fieldAlias:
 					Visit(fieldAlias, propertyInfo, attribute);
 					break;
+				case IKnnVectorProperty knnVector:
+					Visit(knnVector, propertyInfo, attribute);
+					break;
 			}
 		}
 	}

src/OpenSearch.Client/QueryDsl/Abstractions/Container/IQueryContainer.cs

@@ -200,6 +200,9 @@ public interface IQueryContainer
 		[DataMember(Name = "distance_feature")]
 		IDistanceFeatureQuery DistanceFeature { get; set; }
 
+		[DataMember(Name = "knn")]
+		IKnnQuery Knn { get; set; }
+
 		void Accept(IQueryVisitor visitor);
 	}
 }

src/OpenSearch.Client/QueryDsl/Abstractions/Container/QueryContainer-Assignments.cs

@@ -51,6 +51,7 @@ public partial class QueryContainer : IQueryContainer, IDescriptor
 		private IHasParentQuery _hasParent;
 		private IIdsQuery _ids;
 		private IIntervalsQuery _intervals;
+		private IKnnQuery _knn;
 		private IMatchQuery _match;
 		private IMatchAllQuery _matchAllQuery;
 		private IMatchBoolPrefixQuery _matchBoolPrefixQuery;
@@ -193,6 +194,12 @@ IIntervalsQuery IQueryContainer.Intervals
 			set => _intervals = Set(value);
 		}
 
+		IKnnQuery IQueryContainer.Knn
+		{
+			get => _knn;
+			set => _knn = Set(value);
+		}
+
 		IMatchQuery IQueryContainer.Match
 		{
 			get => _match;

src/OpenSearch.Client/QueryDsl/Abstractions/Container/QueryContainerDescriptor.cs

@@ -237,6 +237,9 @@ public QueryContainer HasChild<TChild>(Func<HasChildQueryDescriptor<TChild>, IHa
 		public QueryContainer HasParent<TParent>(Func<HasParentQueryDescriptor<TParent>, IHasParentQuery> selector) where TParent : class =>
 			WrapInContainer(selector, (query, container) => container.HasParent = query);
 
+		public QueryContainer Knn(Func<KnnQueryDescriptor<T>, IKnnQuery> selector) =>
+			WrapInContainer(selector, (query, container) => container.Knn = query);
+
 		/// <summary>
 		/// A query that generates the union of documents produced by its subqueries, and that scores each document
 		/// with the maximum score for that document as produced by any subquery, plus a tie breaking increment for

src/OpenSearch.Client/QueryDsl/Query.cs

@@ -92,6 +92,9 @@ public static QueryContainer Ids(Func<IdsQueryDescriptor, IIdsQuery> selector) =
 		public static QueryContainer Intervals(Func<IntervalsQueryDescriptor<T>, IIntervalsQuery> selector) =>
 			new QueryContainerDescriptor<T>().Intervals(selector);
 
+		public static QueryContainer Knn(Func<KnnQueryDescriptor<T>, IKnnQuery> selector) =>
+			new QueryContainerDescriptor<T>().Knn(selector);
+
 		public static QueryContainer Match(Func<MatchQueryDescriptor<T>, IMatchQuery> selector) =>
 			new QueryContainerDescriptor<T>().Match(selector);