DSpyNet 1.0.0

DSpyNet 🧠

DSpyNet is a C# .NET port of the Stanford DSPy framework.

It allows you to program language models rather than prompt them. Instead of tweaking string prompts manually, you define Signatures (Input/Output contracts) and Modules, and let Optimizers (Teleprompters) automatically tune the prompts and select the best few-shot examples for your specific metrics.

Built on top of Microsoft Semantic Kernel.

🚀 Features & Comparison

DSpyNet adapts the dynamic nature of Python's DSPy to the strongly-typed world of .NET.

📦 Installation

Currently, this is a source-only library. Include the DSpyNet project in your solution.

Dependencies:

• .NET 8.0+
• Microsoft.SemanticKernel
• Microsoft.Extensions.Logging

⚡ Quick Start

1. Define a Signature

Instead of writing a prompt text, define what you need using a C# class.

using DSpyNet.DSPy.Core;

[DspInstruction("Translate the text to the target language.")]
public class TranslationSignature : IDSpySignature
{
    [DspInput(Prefix = "Text to translate:")]
    public string InputText { get; set; }

    [DspInput(Prefix = "Target Language:")]
    public string Language { get; set; }

    [DspOutput(Prefix = "Translation:")]
    public string TranslatedText { get; set; }
}

2. Configure Semantic Kernel

DSpyNet wraps Semantic Kernel to communicate with LLMs.

var builder = Kernel.CreateBuilder();
builder.AddOpenAIChatCompletion("gpt-4o", "YOUR_API_KEY");
var kernel = builder.Build();

// Convert to DSpy ILM
var lm = kernel.ToDSpyLM();

3. Run a Module

Create a predictor based on your signature and run it.

var predictor = new Predict<TranslationSignature>(lm);

var result = await predictor.InvokeAsync(new 
{ 
    InputText = "Hello world", 
    Language = "Spanish" 
});

var prediction = (Prediction)result;
Console.WriteLine(prediction.Get<string>("TranslatedText")); 
// Output: Hola Mundo

🧠 Optimization (Teleprompters)

The power of DSPy is compiling your program to optimize it. The BootstrapFewShot optimizer runs a "Teacher" model over your training data, validates the outputs using your metric, and automatically saves the best examples to the prompt (Few-Shot Learning).

// 1. Define Training Data
var trainset = new List<Example>
{
    Example.From(("Question", "2+2?"), ("Answer", "4")),
    Example.From(("Question", "Capital of France?"), ("Answer", "Paris"))
};

// 2. Define a Metric (Correctness check)
Metric exactMatch = (gold, pred) => 
    gold.Get<string>("Answer") == pred.Get<string>("Answer");

// 3. Setup Modules
var student = new ChainOfThought<QASignature>(lm);

// 4. Compile (Optimize)
var optimizer = new BootstrapFewShot<ChainOfThought<QASignature>>(
    metric: exactMatch, 
    maxBootstrappedDemos: 4
);

// This returns a NEW module with optimized prompts and demos embedded
var compiledProgram = await optimizer.CompileAsync(student, trainset);

// 5. Run Optimized Program
var result = await compiledProgram.InvokeAsync(new { Question = "What is 5 + 5?" });

🏗 Architecture Details

Mutability in a Static Language

In Python, DSPy modifies classes on the fly. In C#, classes are static. DSpyNet solves this by separating Schema (Type) from State (Data).

• Signature (Class): Defines the structure, types, and default instructions (Immutable).
• SignatureState (Object): Holds the actual instruction text and the list of Few-Shot examples (Mutable).

Optimizers (like MIPRO or Bootstrap) clone the Module, modify the SignatureState (changing instructions or adding demos), and return a new instance of the module.

Serialization

You can save optimized modules to disk and load them in production:

// Save optimized state
await compiledProgram.SaveAsync("optimized_math_bot.json");

// Load later
var productionBot = new ChainOfThought<MathSignature>(lm);
await productionBot.LoadAsync("optimized_math_bot.json");

🧪 Integration Tests

The repository includes a RealExampleIntegrationTests project. It contains examples of:

• News Generation: Generating social media posts from raw text.
• Content Guard: Analyzing sentiment and safety using Chain of Thought.
• Intent Classification: Optimizing a classification task using BootstrapFewShot.

To run them, you need to set up your API keys (e.g., OpenAI or RouterAI) in the test base class.

🤝 Contributing

This is an active port. Missing features (ReAct, Code Execution, Advanced Bayesian Optimization) are planned. PRs are welcome!