Models & Pipelines (Advanced)

See also: AI Integration (Learn) for the basics and the ![model] effect.

Status Summary

Sailfin’s AI story is two layers:

Layer	Status	Notes
![model] effect	Enforced today	A language-level capability gate. Any function that calls library code carrying ![model] must declare ![model] in its own effect list; the effect checker rejects the code otherwise.
model / prompt / tool / pipeline block keywords	Removed	These keywords have been removed from the language parser and compiler. AI functionality is being delivered as the sfn/ai library capsule, planned post-1.0.

The ![model] Effect (Enforced)

The ![model] effect is a real, enforced capability. Any function that calls library functions carrying ![model] must declare it, and the effect propagates transitively to callers:

fn ai_call(model_name: string, input: string) -> string ![model] {
    // implemented in sfn/ai capsule (post-1.0)
    return "";
}

fn summarize_invoice(invoice: string) -> string ![model, io] {
    let payload = redact(invoice);
    let summary = ai_call("summarizer", payload);
    print(summary);
    return summary;
}

Removing ![model] from summarize_invoice is a compile-time error — this is the safety property Sailfin ships today.

The sfn/ai Capsule (Planned Post-1.0)

The sfn/ai capsule will provide:

• Model invocation functions carrying ![model] in their signatures
• Typed output schemas and structured response parsing
• Tool dispatch for model-orchestrated function calls
• Provider adapters (OpenAI, Anthropic, and others)

Because these are library functions, not keywords, they can be versioned and updated independently of compiler releases. The ![model] effect gates the entire capsule surface at the language level.

What Is Not Shipped

• model { ... } block declarations — removed from the language
• prompt system { ... } / prompt user { ... } blocks — removed
• tool Name { ... } declarations — removed
• pipeline Name { ... } declarations — removed
• Tensor<...> types, training blocks, PyTorch/Ollama adapters — not implemented

---

The sfn/ai capsule that will host runtime AI behaviour is post-1.0 work. See the roadmap.