ai and typed output

ai is a keyword. Nothing to import, nothing to configure.

answer = ai "What is the capital of France?"

An untyped call returns text. The interesting form asks for a type:

count = ai[Int] "How many words in: {text}"
print(count + 1)              # arithmetic, not string concatenation

ai[Int] is an Int. Not "seven", not " 7\n", not a JSON blob you have to parse. The type is enforced by generating a schema from it, sending that schema to the provider, and coercing the reply through it - and if coercion fails, the call retries with the failure fed back before it gives up.

The types you can ask for

ai[Int] "how many?"
ai[Float] "what fraction?"
ai[Bool] "is this spam?"
ai[String] "rewrite this"

type Sentiment = Positive | Negative | Neutral
ai[Sentiment] "classify this review"          # one of the variants

type Person = { name: String, age: Int }
ai[Person] "extract the person from: {text}"  # a struct, fields coerced

ai[json<List<Int>>] "list three primes"       # a shaped JSON value

An enum returns a variant, so it slots straight into match with exhaustiveness checked at analysis time. A struct returns a struct with each field coerced to its declared type.

Why a keyword and not a library

Because the alternative is worse in a specific way. A library call returns a string and hands you the problem: parse it, validate it, decide what to do when it is not what you asked for, remember to do all three at every call site. That scaffolding is where AI programs actually break, and no amount of library design removes it, because the language cannot see inside the string.

Making the model call a language construct means the schema, the coercion, the retry, the budget, the trace and the contract all live in one place - and the type is checked before your next line runs.

Pipelines

ai composes with |>, which feeds the left value in as prompt input:

dates = text
    |> ai "Extract all dates"
    |> ai "Format as ISO 8601"

Clauses

The ai expression takes modifiers, each documented on its own page:

Not every combination is legal - voting, tools, sessions and streaming are mutually exclusive in the current version, and an illegal combination is a clear parse error rather than a surprise at runtime.

No key required

Everything above runs offline. Without a provider, typed calls return deterministic schema-valid values - see Mock mode.