Get Started

AI Constructs

Sailfin treats AI as a first-class concern with dedicated syntax for models, prompts, pipelines, and tools.

Models

Declare model bindings with provider and version:

model gpt4o = openai:"gpt-4o";
model claude = anthropic:"claude-sonnet-4-20250514";

Prompts

Type-safe prompt composition with interpolation:

fn summarize(text: String) -> String ![model] {
    let result = prompt gpt4o ![model] {
        system "You are a concise summarizer. Respond in 2-3 sentences."
        user "Summarize the following:\n\n{{text}}"
    };
    return result;
}

Pipelines

Chain operations into reusable pipelines:

pipeline analyze_feedback {
    step classify(input: String) -> Category ![model] {
        return prompt gpt4o ![model] {
            system "Classify feedback as: positive, negative, or neutral."
            user "{{input}}"
        };
    }

    step extract_themes(input: String) -> Array<String> ![model] {
        return prompt gpt4o ![model] {
            system "Extract key themes as a list."
            user "{{input}}"
        };
    }
}

Tools

Expose functions as tools for model invocation:

tool lookup_user {
    description "Look up a user by ID"
    param id: String "The user ID to look up"

    fn execute(id: String) -> User ![io] {
        return db.find_user(id);
    }
}

Generation Cards

Every model call produces a generation card with provenance metadata:

  • Input hash
  • Model version
  • Cost and latency
  • Random seed (for reproducibility)

Next Steps