Jun 2025 (v0.3.0)

Mochi v0.3.0 introduces first-class generative AI support through the new generate text block. Programs can now invoke large language models (LLMs) directly from within Mochi code, using natural language prompts to generate text at runtime.

This release focuses on providing a clean, testable, and configurable interface for AI-driven code generation, while preserving the simplicity and composability of the language.

New Language Feature: generate text

The generate text block allows you to call out to an LLM using a single prompt and receive a string result. The block behaves like any other expression in the language and can be composed in assignments, functions, or output operations.

Example

let poem = generate text {
  prompt: "Write a haiku about spring"
}
print(poem)

The prompt must be a string. The result is a string produced by the model in response to the prompt.

Prompt Composition

You can build prompts dynamically using string expressions:

let topic = "rain"
let result = generate text {
  prompt: "Describe the feeling of " + topic
}

This enables programmatic generation of content using user input, time of day, or any other computed values.

Runtime Behavior

The generate block uses a default LLM client configured via environment variables. This allows you to plug in any compatible backend without changing code.

Environment Configuration

Variable	Description
LLM_PROVIDER	Provider name (e.g., openai, echo, local)
LLM_DSN	Connection string or base URL with credentials
LLM_MODEL	Default model name (e.g., gpt-4, mistral)

If not set, LLM_PROVIDER defaults to "echo" (which returns the prompt as the output). This is useful for testing.

Example

LLM_PROVIDER=openai \
LLM_DSN='https://api.openai.com/v1?api_key=sk-...' \
LLM_MODEL=gpt-4 \
mochi run hello.mochi

Mochi initializes the default LLM client lazily. If a generate block is evaluated and no client has been configured, it attempts to use the environment to open a connection. Errors are reported at runtime if this fails.

Testing and Reproducibility

To support reproducible tests, the generate block integrates with the test framework. Generated output can be snapshot-tested using expect blocks:

test "generate greeting" {
  let result = generate text {
    prompt: "Say hello in Japanese"
  }

  expect result == "こんにちは"
}

Run tests with mochi test --update to regenerate outputs. This ensures consistency across test runs and allows AI behavior to be versioned alongside code.