Skip to content

AI Integration

Sailfin gates AI operations through the ![model] effect — the same capability system used for IO, networking, and other side effects. The model, prompt, pipeline, and tool block keywords have been removed from the language. AI functionality will be delivered as a library via the sfn/ai capsule, planned for post-1.0. The ![model] effect remains as the only AI-specific language feature.

Why a Library, Not Language Syntax

Section titled “Why a Library, Not Language Syntax”

AI APIs change faster than compiler release cycles. Embedding model providers, prompt formats, and orchestration patterns as language syntax would make the grammar a moving target. Delivering them as a library capsule allows independent iteration without breaking existing code.

The capability gate — ![model] — stays in the language because it is a compile-time security property, not an API shape. Any function that calls into AI infrastructure must declare this effect, so the capability surface remains auditable regardless of which library version is in use.

The ![model] Effect

Section titled “The ![model] Effect”

The ![model] effect marks any function that calls into the sfn/ai capsule or any other library function carrying ![model] in its signature. Callers are expected to declare the same effect; full call-graph–transitive enforcement based on callee signatures is planned but not yet implemented.

// A library function that carries ![model] (e.g., from sfn/ai post-1.0)
fn ai_call(model_name: string, input: string) -> string ![model] {
    // implementation in sfn/ai capsule
}


// Callers must also declare ![model]
fn summarize(text: string) -> string ![model] {
    return ai_call("summarizer", text);
}


// Callers of summarize must also declare ![model]
fn process_document(doc: string) -> string ![io, model] {
    let summary = summarize(doc);
    print("Done");
    return summary;
}

Omitting ![model] from a function that directly uses a model-capable API is a compile error. Transitive enforcement across callers is tracked under Effect System Hardening on the roadmap and will land alongside the sfn/ai capsule.

Roadmap

Section titled “Roadmap”

The sfn/ai capsule is planned for post-1.0. It will provide model invocation, typed output schemas, and tool dispatch as library functions. See the roadmap for the current timeline.

Next Steps

Section titled “Next Steps”