rust-ringo-fm-sdk

Rust bindings for Apple's on-device Foundation Models framework.

Workspace layout

• crates/ringo-fm-sys — raw FFI bindings (bindgen) to the Swift bridge from ringo-fm-bridge.
• crates/ringo-fm — idiomatic, async (tokio) wrapper. Public surface mirrors the Python package: SystemLanguageModel, LanguageModelSession, Prompt, GenerationSchema, Tool, Transcript, …
• crates/ringo-fm/examples/ — Rust ports of the three Python examples.

Requirements

• macOS 26+ with Apple Intelligence enabled.
• Full Xcode 26+ install (xcode-select pointing at Xcode.app, not just the CLI tools).
• The vendored crates/ringo-fm-sys/vendor/foundation-models-c Swift package included in this repository, or set APPLE_FM_SDK_SWIFT_PKG to another compatible foundation-models-c/ directory.

Quick start

use ringo_fm::{LanguageModelSession, SystemLanguageModel};

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let model = SystemLanguageModel::default()?;
    let (ok, reason) = model.is_available();
    if !ok { eprintln!("model unavailable: {reason:?}"); return Ok(()); }

    let session = LanguageModelSession::new(
        Some(&model),
        Some("You are concise."),
        Vec::new(),
    )?;
    let response = session.respond("What is the capital of France?").await?;
    println!("{response}");
    Ok(())
}

Build & run

cd crates/ringo-fm-sys/vendor/foundation-models-c
swift build -c release

cd ../../../..
cargo build --workspace
cargo run --example simple_inference
cargo run --example streaming_example
cargo run --example transcript_processing -- path/to/transcript.json

The Rust examples automatically add an rpath to crates/ringo-fm-sys/vendor/foundation-models-c/.build/release. If your Swift package checkout lives elsewhere, set APPLE_FM_SDK_SWIFT_PKG to the absolute path of that foundation-models-c directory before building or running.

Schema Discovery

LanguageModelSession::discover_schema uses guided generation to infer a reviewable schema candidate from text, JSON, or already-extracted documents. The response includes field candidates, evidence, warnings, metrics, and review findings; treat it as a draft for human review, not an approved schema.

use ringo_fm::{
    DiscoveryDocument, DiscoveryDocumentSource, DiscoverSchemaRequest,
    DiscoveryOptions, GenerationOptions, LanguageModelSession,
};

# async fn example() -> ringo_fm::Result<()> {
let session = LanguageModelSession::default()?;
let response = session.discover_schema(
    DiscoverSchemaRequest {
        documents: vec![DiscoveryDocument {
            id: "doc-1".into(),
            source: DiscoveryDocumentSource {
                source_type: "text".into(),
                content: Some("請求日 2026-01-01\n合計 12,000円".into()),
                ..Default::default()
            },
            metadata: None,
        }],
        options: Some(DiscoveryOptions::default()),
        ..Default::default()
    },
    &GenerationOptions::default(),
).await?;
# let _ = response;
# Ok(())
# }

Status

V1 ships everything from the Python public API except:

• #[derive(Generable)] proc macro — implement Generable manually for now.
• Multi-tool sessions — the C bridge maps every tool call through a single trampoline symbol, so v1 reliably supports one registered tool per session. Multi-tool dispatch needs a libffi-style closure trampoline; tracked as a follow-up.

License

Apache-2.0

Third-party code

This repository vendors crates/ringo-fm-sys/vendor/foundation-models-c from ringo-fm-bridge, which is licensed under Apache-2.0. The vendored package retains its original source headers, and a copy of the bridge license is included at crates/ringo-fm-sys/vendor/foundation-models-c/LICENSE.md.