Type System

ThinkLang has a structural type system designed for AI-generated data. Types define the shape that AI outputs must conform to.

Primitive Types

Type	Description	Example values
string	Text	"hello"
int	Integer	42
float	Floating-point number	3.14
bool	Boolean	true, false
null	Null value	null

Named Types

Define structured types with type:

type Person {
  name: string
  age: int
  email: string
}

Types compile to JSON schemas that guide the LLM's structured output. The AI is constrained to return data matching your type exactly.

Field Annotations

Add metadata to fields with annotations. These guide the AI's output:

type Summary {
  @description("One-line headline for the article")
  headline: string
  @description("Key takeaways, maximum 5")
  @maxItems(10)
  keyPoints: string[]
  @description("Overall tone: positive, negative, or neutral")
  tone: string
}

Arrays

Use T[] for arrays of any type:

type SearchResult {
  results: string[]
  scores: float[]
  people: Person[]
}

Optional Types

Use T? for fields that may be absent:

type Person {
  name: string
  role: string
  company: string?
}

Optional fields are not required in the AI's response.

Union Types

Use A | B for values that can be one of several types:

type Response {
  status: string
  data: string | int | null
}

Confident<T>

Wrap any type in Confident<T> to request confidence metadata from the AI:

type Category {
  name: string
  subcategory: string
}

let result = think<Confident<Category>>("Classify this product")
  with context: item

The AI returns an object with value (your data), confidence (0 to 1), and reasoning (explanation). See Confidence for details.

Type Annotations on Variables

You can annotate variable types explicitly:

let name: string = "Alice"
let scores: float[] = [9.5, 8.2, 7.1]

In most cases, types are inferred from the expression on the right-hand side.

How Types Map to AI Output

When you write think<MyType>(...), ThinkLang:

1. Converts MyType to a JSON schema
2. Sends the schema to the LLM as a structured output constraint
3. The LLM returns data conforming to that schema
4. ThinkLang validates and returns the typed result

This means type definitions directly control what the AI produces. Be precise in your type names, field names, and annotations -- they all influence the quality of AI output.