Implemented very first vesrion of hyde.

Author: alkampfergit

src/KernelMemory.Extensions.ConsoleTest/Samples/CustomSearchPipelineBase.cs

@@ -27,25 +27,28 @@ public async Task RunSample2()
     {
         var services = new ServiceCollection();
 
-        var apiKey = Dotenv.Get("COHERE_API_KEY")!;
-        var cohereBaseUrl = Dotenv.Get("COHERE_BASE_API_KEY");
-        if (string.IsNullOrEmpty(cohereBaseUrl))
+        var cohereAzureBaseUrl = Dotenv.Get("COHERE_AZURE_BASE_URL");
+        if (string.IsNullOrEmpty(cohereAzureBaseUrl))
         {
+            var apiKey = Dotenv.Get("COHERE_API_KEY")!;
             services.ConfigureCohereChat(apiKey);
         }
         else
         {
-            services.ConfigureCohereChat(apiKey, cohereBaseUrl);
+            var azureApiKey = Dotenv.Get("COHERE_AZURE_API_KEY");
+            services.ConfigureCohereChat(azureApiKey, cohereAzureBaseUrl);
         }
         //verify if rerank has a different api key (because the apikey point on azure ai studio)
-        var rerankApiKey = Dotenv.Get("COHERE_RERANK_API_KEY");
-        if (string.IsNullOrEmpty(rerankApiKey))
+        var rerankAzureBaseUrl = Dotenv.Get("COHERE_AZURE_RERANK_BASE_URL");
+        if (string.IsNullOrEmpty(rerankAzureBaseUrl))
         {
+            var apiKey = Dotenv.Get("COHERE_API_KEY")!;
             services.ConfigureCohereRerank(apiKey);
         }
         else
         {
-            services.ConfigureCohereRerank(rerankApiKey);
+            var azureReRankApiKey = Dotenv.Get("COHERE_AZURE_RERANK_API_KEY");
+            services.ConfigureCohereRerank(azureReRankApiKey, rerankAzureBaseUrl);
         }
 
         services.AddHttpClient<RawCohereChatClient>()
@@ -89,12 +92,15 @@ public async Task RunSample2()
             .Title("Select query rewriter")
             .AddChoices(["Semantic Kernel Base", "Semantic Kernel Handlebar"]));
 
+        var useHyde = AnsiConsole.Confirm("Do you want to use HyDe? (y/n)", false);
+
         var kernelBuider = CreateBasicKernelBuilder();
         var builder = CreateBasicKernelMemoryBuilder(
             services,
             storageToUse == "elasticsearch",
             queryExecutorToUse,
-            queryRewriterTool == "Semantic Kernel Handlebar");
+            queryRewriterTool == "Semantic Kernel Handlebar",
+            useHyde);
         var kernelMemory = builder.Build<MemoryServerless>();
         var kernel = kernelBuider.Build();
 
@@ -238,7 +244,8 @@ private static IKernelMemoryBuilder CreateBasicKernelMemoryBuilder(
         ServiceCollection services,
         bool useElasticSearch,
         string ragToolToUse,
-        bool useHandlebarQueryRewriter)
+        bool useHandlebarQueryRewriter,
+        bool useHyde)
     {
         // we need a series of services to use Kernel Memory, the first one is
         // an embedding service that will be used to create dense vector for
@@ -306,6 +313,12 @@ private static IKernelMemoryBuilder CreateBasicKernelMemoryBuilder(
         services.AddSingleton<HandlebarSemanticKernelQueryRewriter>();
         services.AddSingleton<SemanticKernelQueryRewriter>();
         services.AddSingleton<StandardVectorSearchQueryHandler>();
+        services.AddSingleton<HyDeQueryHandler>();
+        var hydeConfig = new HiDeQueryHandlerConfiguration()
+        {
+            Prompt = "Given a question, generate a paragraph of text that answers the question in the context of computer security and IT security"
+        };
+        services.AddSingleton(hydeConfig);
         services.AddSingleton<KeywordSearchQueryHandler>();
 
         var rewriterOptions = new SemanticKernelQueryRewriterOptions();
@@ -337,6 +350,11 @@ private static IKernelMemoryBuilder CreateBasicKernelMemoryBuilder(
                 config.AddHandler<KeywordSearchQueryHandler>();
             }
 
+            if (useHyde)
+            {
+                config.AddHandler<HyDeQueryHandler>();
+            }
+
             if (ragToolToUse == "Cohere CommandR+")
             {
                 config.AddHandler<CohereCommandRQueryExecutor>();

src/KernelMemory.Extensions/QueryPipeline/HyDeQueryHandler.cs

@@ -0,0 +1,95 @@
+using DocumentFormat.OpenXml.Packaging;
+using Microsoft.Extensions.Logging;
+using Microsoft.KernelMemory.Diagnostics;
+using Microsoft.KernelMemory.MemoryStorage;
+using Microsoft.SemanticKernel;
+using System.Collections.Generic;
+using System.Text;
+using System.Threading;
+using System.Threading.Tasks;
+
+namespace KernelMemory.Extensions;
+
+public class HiDeQueryHandlerConfiguration 
+{
+    public string Prompt { get; set; } = "Given a question, generate a paragraph of text that answers the question.";
+}
+
+public class HyDeQueryHandler : BasicQueryHandler
+{
+    private readonly IMemoryDb _memoryDb;
+    private readonly Kernel _kernel;
+    private readonly HiDeQueryHandlerConfiguration _configuration;
+    private readonly ILogger<StandardVectorSearchQueryHandler> _log;
+
+    public override string Name => "StandardVectorSearchQueryHandler";
+
+    public HyDeQueryHandler(
+        IMemoryDb memory,
+        Kernel kernel,
+        HiDeQueryHandlerConfiguration? configuration,
+        ILogger<StandardVectorSearchQueryHandler>? log = null)
+    {
+        _memoryDb = memory;
+        _kernel = kernel;
+        _configuration = configuration ?? new HiDeQueryHandlerConfiguration();
+        _log = log ?? DefaultLogger<StandardVectorSearchQueryHandler>.Instance;
+    }
+
+    /// <summary>
+    /// Perform a vector search in default memory using the hyde principle.
+    /// </summary>
+    /// <param name="userQuestion"></param>
+    /// <param name="cancellationToken"></param>
+    /// <returns></returns>
+    protected override async Task OnHandleAsync(UserQuestion userQuestion, CancellationToken cancellationToken)
+    {
+        // Perform a vector search in default memory
+        StringBuilder prompt = new StringBuilder(_configuration.Prompt.Length + userQuestion.Question.Length + "Question: ".Length + "Paragraph: ".Length + 20);
+        prompt.AppendLine(_configuration.Prompt);
+        prompt.AppendLine("Question: " + userQuestion.Question);
+        prompt.AppendLine("Paragraph: ");
+
+        var result = await _kernel.InvokePromptAsync(prompt.ToString(), cancellationToken: cancellationToken);
+
+        var paragraph = result.ToString();
+
+        var list = new List<(MemoryRecord memory, double relevance)>();
+
+        IAsyncEnumerable<(MemoryRecord, double)> matches = this._memoryDb.GetSimilarListAsync(
+            index: userQuestion.UserQueryOptions.Index,
+            text: paragraph,
+            filters: userQuestion.Filters,
+            minRelevance: userQuestion.UserQueryOptions.MinRelevance,
+            limit: userQuestion.UserQueryOptions.RetrievalQueryLimit,
+            withEmbeddings: false,
+            cancellationToken: cancellationToken);
+
+        // Memories are sorted by relevance, starting from the most relevant
+        await foreach ((MemoryRecord memory, double relevance) in matches.ConfigureAwait(false))
+        {
+            list.Add((memory, relevance));
+        }
+
+        var records = new List<MemoryRecord>();
+        // Memories are sorted by relevance, starting from the most relevant
+        foreach ((MemoryRecord memory, double relevance) in list)
+        {
+            var partitionText = memory.GetPartitionText(this._log).Trim();
+            if (string.IsNullOrEmpty(partitionText))
+            {
+                this._log.LogError("The document partition is empty, doc: {0}", memory.Id);
+                continue;
+            }
+
+            if (relevance > float.MinValue)
+            {
+                this._log.LogTrace("Adding result with relevance {0}", relevance);
+                records.Add(memory);
+            }
+        }
+
+        //ok now that you have all the memory record and citations, add to the object
+        userQuestion.AddMemoryRecordSource("hyde-vector-search", records);
+    }
+}