Jun 2025 (v0.3.1)

Mochi v0.3.1 expands on the generative AI foundation introduced in v0.3.0 by adding first-class support for user-defined types and structured AI output generation. Programs can now declare their own rich types, instantiate them manually, or populate them dynamically using AI-generated responses.

This release further emphasizes Mochi’s commitment to clarity, composability, and deep integration with modern AI workflows.

New Language Feature: User-Defined Types

You can now declare custom struct-like types using the type keyword. These types group related fields under a named structure, enabling safer and more expressive programs.

Example

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

type Book {
  title: string
  author: Person
  pages: int
  tags: list<string>
  metadata: map<string, string>
  published: bool
}

Usage

Declared types can be used to declare variables, construct values, and access fields:

let bob: Person = Person {
  name: "Bob",
  age: 42,
  email: "[email protected]"
}

let book: Book = Book {
  title: "Example",
  author: bob,
  pages: 100,
  tags: ["guide"],
  metadata: { "isbn": "123" },
  published: true
}

print("custom types defined")
print(book.author.name)

Features

• ✅ Full support for nested structs, lists, and maps
• ✅ Inline value construction using { field: value, ... } syntax
• ✅ Type-annotated variables (let x: Person)
• ✅ Field access with dot notation (x.field)

This forms the foundation for composable data modeling, JSON-like transformations, and generative type synthesis.

New Language Feature: generate <Type>

Building on user-defined types, you can now use LLMs to synthesize structured data directly:

let p = generate Person {
  prompt: "Generate a fictional software engineer with name, age, and email."
}

Mochi parses the returned JSON and validates it against the declared Person type. All fields must be present and match their expected types.

Example

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

let p = generate Person {
  prompt: "Generate a fictional software engineer"
}

print("Name:",  p.name)
print("Age:",   p.age)
print("Email:", p.email)

Error Handling

Mochi validates the LLM’s response against your type. If a required field is missing or the structure doesn't match, you’ll receive a detailed runtime error with a diff.

Prompt Composition and Functions

All prompt text is fully dynamic:

fun describe(topic: string): Book {
  generate Book {
    prompt: "Generate a book about " + topic
  }
}

Runtime Configuration

As in v0.3.0, the LLM backend is defined via environment variables:

Variable	Description
LLM_PROVIDER	Provider name (e.g., openai, echo, local)
LLM_DSN	Connection string or API base URL
LLM_MODEL	Default model name

Structured and text generation both use this shared configuration.

Testing and Snapshot Support

generate blocks integrate with Mochi’s test system. Use expect assertions to snapshot structured output:

test "generate book" {
  let book = generate Book {
    prompt: "Create a guidebook named 'Mochi Language'"
  }

  expect book.title == "Mochi Language"
}

Run mochi test --update to re-record the generated output.

Mochi v0.3.1 makes typed programming and generative AI first-class citizens — setting the stage for upcoming agent, stream, and task modeling features in the next release.