Implementing Semantic Kernel with ASP.NET Core 8: A Reference Architecture

The Paradigm Shift in Enterprise AI

Integrating Large Language Models (LLMs) into an enterprise environment requires far more than just passing a string to the OpenAI API. It requires orchestration, prompt management, secure memory injection (RAG), and strict multi-tenant boundary enforcement. This is where Microsoft's Semantic Kernel shines.

Why Semantic Kernel over LangChain?

While LangChain dominates the Python ecosystem, enterprise environments heavily invested in C# and .NET need a native, strongly-typed orchestrator. Semantic Kernel provides a highly abstraction-oriented approach to AI plugins, allowing us to swap models (e.g., from OpenAI GPT-4 to a local Llama 3 instance) without rewriting business logic.

The Architecture

In our reference architecture, we inject the Kernel as a scoped service in the ASP.NET Core Dependency Injection container. Crucially, we configure the kernel dynamically per-request using a custom ITenantModelProvider.

// Example: Dynamic Kernel Injection
builder.Services.AddScoped<Kernel>(sp => {
    var tenantConfig = sp.GetRequiredService<ITenantModelProvider>().GetCurrentConfig();
    var builder = Kernel.CreateBuilder();
    builder.AddOpenAIChatCompletion(tenantConfig.ModelId, tenantConfig.ApiKey);
    return builder.Build();
});

Security and Multi-Tenancy

By scoping the Kernel to the DI container's request lifetime, we guarantee that API keys and context windows do not bleed across tenants. If Tenant A is using Azure OpenAI and Tenant B is using Anthropic, the pipeline dynamically routes the execution at runtime.

Conclusion

Semantic Kernel provides the necessary guardrails for enterprise AI. By treating AI models as interchangeable dependencies, we achieve maximum flexibility while maintaining strict security boundaries.