Five minutes from zero to running.

# Create a new console project.
dotnet new console -n MyFirstLMKit
cd MyFirstLMKit

# Add the LM-Kit package.
dotnet add package LM-Kit.NET

# Optional: target .NET 8+ explicitly in the .csproj if you need to.
# <TargetFramework>net8.0</TargetFramework>

using LMKit.Model;
using LMKit.TextGeneration;

var model = LM.LoadFromModelID("qwen3.5:4b");
var chat  = new MultiTurnConversation(model)
{
    SystemMessage = "You are a concise expert in .NET and AI."
};

// Stream the response token-by-token.
await foreach (var token in chat.StreamAsync("What is a KV-cache?"))
{
    Console.Write(token.Text);
}
Console.WriteLine();

// Follow-up. The model has the previous turn in context.
var followUp = await chat.SubmitAsync("How does it relate to context length?");
Console.WriteLine(followUp);

using LMKit.Model;
using LMKit.Retrieval;
using LMKit.Data.Storage;
using LMKit.TextGeneration;

// 1. Models. Chat for answers, embeddings for retrieval.
var chatModel  = LM.LoadFromModelID("qwen3.5:4b");
var embedModel = LM.LoadFromModelID("embeddinggemma-300m");

// 2. RAG engine. File-based vector store, persists across runs.
var store = new FileSystemVectorStore("./embeddings");
var rag   = new DocumentRag(embedModel, store)
{
    ProcessingMode = PageProcessingMode.Auto,
    MaxChunkSize   = 512
};

// 3. Import a PDF. Ingestion handles text-layer pages and scans automatically.
await rag.ImportDocumentAsync(
    Attachment.FromFile("sample.pdf"),
    new("sample") { Name = "Sample document" });

// 4. Query. The response includes the source pages it cited.
var chat   = new SingleTurnConversation(chatModel);
var result = await rag.QueryPartitionsAsync(
    "Summarise the document in three sentences.",
    chat);

Console.WriteLine(result.Response.Completion);
foreach (var r in result.SourceReferences)
{
    Console.WriteLine($"  cited: {r.Name}, page {r.PageNumber}");
}