Struct EdgeLlm 

pub struct EdgeLlm { /* private fields */ }

The flat FFI-friendly facade (ADR-001, ADR-009, ADR-010).

Annotated for all three binding surfaces:

• uniffi::Object (native) → opaque UniFFI / React Native handle
• frb(opaque) (native) → opaque Dart handle via FRB v2 codegen
• wasm_bindgen (wasm32) → satisfies IntoWasmAbi/WasmDescribe so that #[wasm_bindgen] impl EdgeLlm { ... } compiles

Implementations

impl EdgeLlm

UniFFI-exported methods: constructors, blocking chat, and reset.

Dart uses the edge_llm_* free-function wrappers below so FRB does not need to parse the UniFFI-decorated impl block after macro expansion.

pub fn local(model_uri: String) -> Result<Self, SdkError>

Construct with the local Candle engine (air-gapped, ADR-002/004).

If model_uri is non-empty, loads the GGUF at that path via CandleEngine::from_path (consumer-supplied model, ADR-002). Pass an empty string to use a deterministic toy model for development and testing — the toy generates gibberish but exercises the full binding layer end-to-end.

Returns Err(SdkError) if model_uri is non-empty but the file cannot be parsed (missing, malformed GGUF, incompatible tensor shapes). An empty model_uri never fails.

The permissive signature verifier used here is intentional: it lets the binding layer be exercised without a signed model artifact. A production deployment should substitute a real SignatureVerifier backed by the platform keystore.

pub fn cloud(model: String, api_key: String) -> Self

Construct with a frontier cloud backend (opt-in, ADR-010).

model uses the routing prefix: "openai/gpt-4o", "anthropic/claude-sonnet-4-6", "ollama/llama3", "gemini/gemini-2.0-flash", or any OpenAI-compat base URL. api_key must come from the platform keystore — never embedded.

Native only (React Native / Dart native). On wasm32 this constructor does not exist — the web surface exposes a throwing cloud instead (see the wasm32 impl block below and the ADR-010 amendment).

pub fn ask(&self, prompt: String) -> Result<String, SdkError>

Blocking chat completion.

Returns Err(SdkError::ProviderError) on network/auth/engine failure so callers can distinguish model output from error conditions.

pub fn reset(&self)

Reset the session (clears KV cache and output).

impl EdgeLlm

Streaming via callback interface — exported for React Native (UniFFI).

Separated from the main block because UniFFI cannot export impl FnMut.

pub fn ask_stream_cb( &self, prompt: String, handler: Box<dyn StreamHandler>, ) -> Result<(), SdkError>

Stream tokens to a StreamHandler callback (React Native path).

Returns an error on network/auth/engine failure so callers are not left waiting for a stream that will never arrive.

Trait Implementations

impl<UT> LiftRef<UT> for EdgeLlm

type LiftType = Arc<EdgeLlm>

impl<UT> LowerError<UT> for EdgeLlm

fn lower_error(obj: Self) -> RustBuffer

Lower this value for scaffolding function return

impl<UT> LowerReturn<UT> for EdgeLlm

type ReturnType = <Arc<EdgeLlm> as LowerReturn<UniFfiTag>>::ReturnType

The type that should be returned by scaffolding functions for this type.

fn lower_return(obj: Self) -> Result<Self::ReturnType, RustCallError>

Lower the return value from an scaffolding call

fn handle_failed_lift( error: LiftArgsError, ) -> Result<Self::ReturnType, RustCallError>

Lower the return value for failed argument lifts

impl<UT> TypeId<UT> for EdgeLlm

const TYPE_ID_META: MetadataBuffer