Structured Output

Why Structured Output?

LLMs return free-form text by default. If your application needs to process the response programmatically — store it in a database, pass it to another API, or render it in a typed UI — you have to parse that text yourself. This leads to fragile regex or string.Contains checks that break when the model changes phrasing.

Structured output solves this by instructing the model to return JSON matching a C# type's schema. Mythosia.AI handles the schema generation, prompt injection, and deserialization automatically — including automatic JSON repair for minor formatting errors the model may produce.

When to Use

Extracting entities, classifications, or structured data from unstructured text
Building typed API responses from AI-generated content
Feeding AI output into downstream pipelines that expect specific data shapes
Any scenario where you need reliable, machine-readable output from the model

The Problem It Solves

Suppose you need to extract weather data from the model's response. Without structured output:

// ❌ Without structured output — fragile manual parsing
var text = await service.GetCompletionAsync("What's the weather in Seoul?");
// text = "The weather in Seoul is sunny with a temperature of 22°C."

// Now you have to parse this yourself...
var city = "Seoul"; // hardcoded? regex?
var tempMatch = Regex.Match(text, @"(\d+)°C");
int temp = tempMatch.Success ? int.Parse(tempMatch.Groups[1].Value) : 0;
// What if the model says "twenty-two degrees" instead of "22°C"? 💥

This breaks whenever the model changes phrasing. With structured output:

// ✅ With structured output — type-safe, automatic
public record WeatherResponse(string City, string Condition, int TemperatureC);

var result = await service.GetCompletionAsync<WeatherResponse>(
    "What's the weather in Seoul?");

Console.WriteLine(result.City);         // Seoul
Console.WriteLine(result.Condition);    // Sunny
Console.WriteLine(result.TemperatureC); // 22

The model is instructed to return JSON matching your C# type. Mythosia.AI deserializes it automatically. If the model produces slightly malformed JSON (missing comma, trailing text), the built-in auto-repair fixes it before deserialization — no manual error handling needed.

Basic

Pass a type parameter to GetCompletionAsync:

public record WeatherResponse(string City, string Condition, int TemperatureC);

var result = await service.GetCompletionAsync<WeatherResponse>(
    "What's the weather in Seoul?");

Console.WriteLine(result.City);        // Seoul
Console.WriteLine(result.Condition);   // Sunny
Console.WriteLine(result.TemperatureC); // 22

Collections

Collection types work directly — no wrapper DTO needed:

public record Entity(string Name, string Type);

var entities = await service.GetCompletionAsync<List<Entity>>(
    "Extract all people and organizations from this text: ...");

foreach (var e in entities)
    Console.WriteLine($"{e.Type}: {e.Name}");

Streaming + Structured Output

Stream text in real-time while also getting the final deserialized object:

var run = service.BeginStream("Generate a product summary").As<ProductDto>();

// Real-time output
await foreach (var chunk in run.Stream())
    Console.Write(chunk);

// Final parsed result
ProductDto product = await run.Result;