Add Microsoft Agent Framework (Microsoft.Agents.AI) with RAG Example

LLama.Examples/ExampleRunner.cs

@@ -23,6 +23,7 @@ public class ExampleRunner
         { "LLama Model: Get embeddings", GetEmbeddings.Run },
         { "LLama Model: Quantize", QuantizeModel.Run },
         { "Grammar: Constrain response to json format", GrammarJsonResponse.Run },
+        { "Microsoft Agent Framework Example", MicrosoftAgentFrameworkExample.Run },
         { "Kernel Memory: Document Q&A", KernelMemory.Run },
         { "Kernel Memory: Save and Load", KernelMemorySaveAndLoad.Run },
 		{ "Semantic Kernel: HomeAutomation", SemanticKernelHomeAutomation.Run },

LLama.Examples/Examples/MicrosoftAgentFrameworkExample.cs

@@ -0,0 +1,181 @@
+/*
+=============================================================================
+# Microsoft Agent Framework & LLamaSharp (with Deterministic RAG)
+
+This example demonstrates how to integrate LLamaSharp with the Microsoft Agent 
+Framework (Microsoft.Agents.AI) and Microsoft's official Vector Data 
+abstractions (Microsoft.Extensions.VectorData).
+
+It implements a Deterministic RAG (Retrieval-Augmented Generation) pipeline. 
+Instead of relying on the LLM to autonomously use tools (which often causes 
+small local models to hallucinate), the C# code explicitly handles the vector 
+search and forces the verified facts into the AI's prompt using a strict 
+persona constraint.
+
+## 🧠 Key Concepts Demonstrated
+* The `IChatClient` Bridge: Wrapping LLamaSharp's `StatelessExecutor` into 
+  Microsoft's standard `ChatClientBuilder`.
+* Embedding Generation: Utilizing LLamaSharp's `LLamaEmbedder` to generate 
+  vectors from text in real-time.
+* Vector Storage Integration: Using `CommunityToolkit.VectorData.SqliteVec` 
+  to power an in-memory SQLite vector database.
+* Strict Agent Constraints: Configuring `ChatClientAgentRunOptions` to lock 
+  the model's Temperature to 0.0f to prevent creative hallucinations.
+
+## 🚀 How to Run the Example
+To run this example successfully, you will need two separate local models:
+1. A Chat Model: Used for the Agent persona (e.g., Llama-3.2-3B-Instruct).
+2. An Embedding Model: Used for Vector Search (e.g., Qwen-Embedding-0.6B).
+
+⚠️ Important Dimension Note:
+By default, this example's `DocumentRecord` class is hardcoded to expect 
+1024 dimensions, which matches the Qwen3 embedding model. If you use a 
+different embedding model (like all-MiniLM-L6-v2), you must change the 
+dimension attribute in the code to match (e.g., 
+`[VectorStoreVector(Dimensions: 384)]`) before compiling, or the SQLite 
+database will throw an error.
+
+## 🔍 How It Works Under the Hood
+1. Ingestion: Embeds dummy strings and upserts them into the SQLite database.
+2. Retrieval: Automatically embeds your query, runs a vector similarity search 
+   against SQLite, and extracts the top 2 matching facts.
+3. Generation: Spins up a Microsoft Agent with strict instructions, feeds it 
+   the facts, and streams a formatted, hallucination-free response.
+4. Cleanup: Because the SQLite database is entirely in-memory, no residual 
+   .db files are left on your hard drive after closing the application.
+=============================================================================
+*/
+
+using CommunityToolkit.VectorData.SqliteVec;
+using LLama.Abstractions;
+using LLama.Common;
+using Microsoft.Agents.AI;
+using Microsoft.Data.Sqlite;
+using Microsoft.Extensions.AI;
+using Microsoft.Extensions.VectorData;
+using Spectre.Console;
+
+namespace LLama.Examples.Examples
+{
+    public class MicrosoftAgentFrameworkExample
+    {
+        public class DocumentRecord
+        {
+            [VectorStoreKey]
+            public long Id { get; set; }
+
+            [VectorStoreData]
+            public string Text { get; set; } = string.Empty;
+
+            // Note: Adjust dimensions to match your embedding model (e.g., 1024 for Qwen3-0.6B)
+            [VectorStoreVector(dimensions: 1024)]
+            public ReadOnlyMemory<float> Vector { get; set; }
+        }
+
+        public static async Task Run()
+        {
+            Console.WriteLine("=== Microsoft Agent Framework & Deterministic RAG ===");
+
+            // 1. Get Model Paths using standard LLama.Examples prompts
+            string chatModelPath = UserSettings.GetModelPath(); //https://huggingface.co/bartowski/Llama-3.2-3B-Instruct-GGUF/resolve/main/Llama-3.2-3B-Instruct-Q4_K_M.gguf?download=true
+            string embedModelPath = UserSettings.GetEmbedModelPath(); //https://huggingface.co/Qwen/Qwen3-Embedding-0.6B-GGUF/blob/main/Qwen3-Embedding-0.6B-Q8_0.gguf
+
+            Console.WriteLine("\nLoading Models into Memory...");
+
+            var chatParams = new ModelParams(chatModelPath) { ContextSize = 4096, GpuLayerCount = -1 };
+            using var chatWeights = LLamaWeights.LoadFromFile(chatParams);
+            var chatExecutor = new StatelessExecutor(chatWeights, chatParams);
+
+            var embedParams = new ModelParams(embedModelPath) { Embeddings = true, GpuLayerCount = -1 };
+            using var embedWeights = LLamaWeights.LoadFromFile(embedParams);
+            using var embedder = new LLamaEmbedder(embedWeights, embedParams);
+
+            // 2. Setup SQLite Vector Database
+            Console.WriteLine("Initializing In-Memory SQLite Vector Database...");
+            string connectionString = "Data Source=SharedVectorDb;Mode=Memory;Cache=Shared";
+
+            // 2a. OPEN A MASTER CONNECTION TO KEEP THE DB ALIVE
+            // This prevents SQLite from destroying the DB between vector store operations
+            using var keepAliveConnection = new SqliteConnection(connectionString);
+            await keepAliveConnection.OpenAsync();
+
+            // 2b. Initialize the vector store using the same connection string
+            var vectorStore = new SqliteVectorStore(connectionString);
+
+            VectorStoreCollection<long, DocumentRecord> documentCollection =
+                vectorStore.GetCollection<long, DocumentRecord>("company_docs");
+
+            await documentCollection.EnsureCollectionExistsAsync();
+
+            // 3. Seed the Database
+            string[] rawDocuments = [
+                "Project Zephyr launches in Q4. The marketing budget is $50,000.",
+                "The CEO announced that Fridays are now remote-work days.",
+                "Project Zephyr's lead engineer is Sarah Connor."
+            ];
+
+            long idCounter = 1;
+            foreach (var doc in rawDocuments)
+            {
+                var embeddingBatch = await embedder.GetEmbeddings(doc);
+                float[] vectorArray = embeddingBatch.First();
+
+                await documentCollection.UpsertAsync(new DocumentRecord
+                {
+                    Id = idCounter++,
+                    Text = doc,
+                    Vector = new ReadOnlyMemory<float>(vectorArray)
+                });
+            }
+
+            // 4. Build the Microsoft Agent ChatClient
+            IChatClient chatClient = new ChatClientBuilder(chatExecutor.AsChatClient()).Build();
+
+            // 5. Run the Interactive Loop
+            Console.WriteLine("\nSetup Complete. Type 'exit' to quit.");
+            while (true)
+            {
+                string task = AnsiConsole.Ask("Enter your research task (or ENTER for default): ", "Draft an email about Project Zephyr");
+                if (task.Equals("exit", StringComparison.OrdinalIgnoreCase))
+                    break;
+
+                Console.WriteLine("\n[System] Searching database...");
+                var embeddingBatch = await embedder.GetEmbeddings(task);
+                float[] queryVector = embeddingBatch.First();
+
+                var searchResults = documentCollection.SearchAsync(
+                    new ReadOnlyMemory<float>(queryVector),
+                    top: 2
+                );
+
+                var retrievedFacts = new List<string>();
+                await foreach (var result in searchResults)
+                {
+                    retrievedFacts.Add($"- {result.Record.Text}");
+                }
+                string contextString = string.Join("\n", retrievedFacts);
+
+                Console.WriteLine($"[System] Facts Retrieved:\n{contextString}\n");
+                Console.WriteLine("[System] Agent is drafting response...");
+
+                var writer = chatClient.AsAIAgent(
+                    name: "Writer",
+                    instructions: "You are a strict corporate communications AI. Your ONLY job is to write a short email using EXACTLY the facts provided by the user. You must NOT add, infer, or invent any outside information."
+                );
+
+                var runOptions = new ChatClientAgentRunOptions
+                {
+                    ChatOptions = new ChatOptions { MaxOutputTokens = 1000, Temperature = 0.0f }
+                };
+
+                string strictPrompt = $"Draft an email to the team using ONLY these exact facts:\n{contextString}\nDo not add any other information.";
+
+                var writerResponse = await writer.RunAsync(strictPrompt, options: runOptions);
+
+                Console.ForegroundColor = ConsoleColor.Green;
+                Console.WriteLine($"\n{writerResponse.Text}\n");
+                Console.ResetColor();
+            }
+        }
+    }
+}

LLama.Examples/LLama.Examples.csproj

@@ -27,6 +27,9 @@
     <PackageReference Include="Whisper.net.Runtime.Clblast" Version="1.5.0" />
     <PackageReference Include="Whisper.net.Runtime.CoreML" Version="1.9.0" />
     <PackageReference Include="Whisper.net.Runtime.Cublas" Version="1.5.0" />
+    <PackageReference Include="Microsoft.Agents.AI" Version="*-*" />
+    <PackageReference Include="Microsoft.Extensions.AI" Version="*-*" />
+    <PackageReference Include="CommunityToolkit.VectorData.SqliteVec" Version="*-*" />
   </ItemGroup>
 
   <ItemGroup>

LLama.Examples/UserSettings.cs

@@ -5,6 +5,7 @@ namespace LLama.Examples;
 internal static class UserSettings
 {
     private static readonly string SettingsModelPath = Path.Join(AppContext.BaseDirectory, "DefaultModel.env");
+    private static readonly string SettingsEmbedModelPath = Path.Join(AppContext.BaseDirectory, "DefaultEmbedModel.env");
     private static readonly string SettingsMMprojPath = Path.Join(AppContext.BaseDirectory, "DefaultMMProj.env");
     private static readonly string SettingsImagePath = Path.Join(AppContext.BaseDirectory, "DefaultImage.env");
     private static readonly string WhisperModelPath = Path.Join(AppContext.BaseDirectory, "DefaultWhisper.env");
@@ -21,6 +22,11 @@ internal static class UserSettings
         return path;
     }
 
+    public static string GetEmbedModelPath(bool alwaysPrompt = false)
+    {
+        return PromptPath("embedmodel.gguf", SettingsEmbedModelPath, alwaysPrompt);
+    }
+
     public static string GetModelPath(bool alwaysPrompt = false)
     {
         return PromptPath("model.gguf", SettingsModelPath, alwaysPrompt);