Models & Pipelines (Advanced)

See also: AI Constructs (Learn) for the basics.

Engine System

Sailfin’s engine system provides a unified interface across AI providers:

model gpt4o = openai:"gpt-4o@2025-09-01";
model bert = torch:"bert@2.5";
model local_llm = ollama:"llama3:8b";

Provider identifiers follow the format: provider:model-name@version

Adapters

Adapters are sandboxed modules implementing the unified model interface. Each provider has an adapter:

• openai — OpenAI API (GPT-4o, o1, etc.)
• anthropic — Anthropic API (Claude)
• torch — PyTorch models
• ollama — Local models via Ollama

Tensors

let weights: Tensor<[768, 512], Float32> = Tensor.zeros();
let input: Tensor<[1, 768], Float32> = encode(text);
let output = input.matmul(weights);  // ![gpu]

Training (Proposed)

training finetune for gpt4o {
    dataset = load_dataset("train.jsonl");
    epochs = 3;
    learning_rate = 1e-5;
}

Generation & Training Cards

Every model interaction produces provenance metadata:

• Generation cards: Input hash, model version, cost, latency, seed
• Training cards: Dataset hash, hyperparameters, metrics, reproducibility info

Capabilities

• ![model] — Required for inference (prompt blocks, model calls)
• ![train] — Required for training operations (proposed)
• PII<T> and Secret<T> enforcement on model inputs

Advanced AI features are specified in the model engines proposal. Implementation is phased across the 1.0 roadmap.