pub trait ModelProvider: Send + Sync {
// Required methods
fn capabilities(&self) -> &Capabilities;
fn chat<'life0, 'async_trait>(
&'life0 self,
request: ChatRequest,
ctx: StreamContext,
) -> Pin<Box<dyn Future<Output = Result<FinalResponse>> + Send + 'async_trait>>
where Self: 'async_trait,
'life0: 'async_trait;
// Provided methods
fn resolve_context_window<'life0, 'life1, 'async_trait>(
&'life0 self,
request: &'life1 ChatRequest,
) -> Pin<Box<dyn Future<Output = ContextSizing> + Send + 'async_trait>>
where Self: 'async_trait,
'life0: 'async_trait,
'life1: 'async_trait { ... }
fn verify_placement<'life0, 'async_trait>(
&'life0 self,
current_num_ctx: Option<usize>,
) -> Pin<Box<dyn Future<Output = Option<ModelPlacement>> + Send + 'async_trait>>
where Self: 'async_trait,
'life0: 'async_trait { ... }
fn supports_vision<'life0, 'async_trait>(
&'life0 self,
) -> Pin<Box<dyn Future<Output = Option<bool>> + Send + 'async_trait>>
where Self: 'async_trait,
'life0: 'async_trait { ... }
}Expand description
Provider-facing interface. A ModelProvider impl owns whatever
HTTP client / state it needs and exposes chat() — that’s the
whole surface.
Required Methods§
Sourcefn capabilities(&self) -> &Capabilities
fn capabilities(&self) -> &Capabilities
Capabilities the provider advertises. The reducer reads this
when building the outgoing ChatRequest (e.g. whether to
attach reasoning controls).
Sourcefn chat<'life0, 'async_trait>(
&'life0 self,
request: ChatRequest,
ctx: StreamContext,
) -> Pin<Box<dyn Future<Output = Result<FinalResponse>> + Send + 'async_trait>>where
Self: 'async_trait,
'life0: 'async_trait,
fn chat<'life0, 'async_trait>(
&'life0 self,
request: ChatRequest,
ctx: StreamContext,
) -> Pin<Box<dyn Future<Output = Result<FinalResponse>> + Send + 'async_trait>>where
Self: 'async_trait,
'life0: 'async_trait,
Stream a chat turn. Typed events flow through
ctx.sink; the returned FinalResponse carries token usage
and the Anthropic thinking-signature (opaque blob required to
continue extended thinking across turns).
Cancellation: the provider MUST select! on ctx.token. cancelled() inside any await that could block for more than
a few hundred ms. This is the contract that replaces the old
check_interrupt polling pattern.
Provided Methods§
Sourcefn resolve_context_window<'life0, 'life1, 'async_trait>(
&'life0 self,
request: &'life1 ChatRequest,
) -> Pin<Box<dyn Future<Output = ContextSizing> + Send + 'async_trait>>where
Self: 'async_trait,
'life0: 'async_trait,
'life1: 'async_trait,
fn resolve_context_window<'life0, 'life1, 'async_trait>(
&'life0 self,
request: &'life1 ChatRequest,
) -> Pin<Box<dyn Future<Output = ContextSizing> + Send + 'async_trait>>where
Self: 'async_trait,
'life0: 'async_trait,
'life1: 'async_trait,
Resolve the effective context window for a turn (what the model will
actually enforce). The default returns the static advertised window;
Ollama overrides this to probe the model’s real window and auto-fit
num_ctx to host memory, honoring the request’s per-model
ollama_num_ctx override. None means “let the backend decide”.
Awaited only on the effect runtime (never the reducer), so a probe never blocks the UI.
Sourcefn verify_placement<'life0, 'async_trait>(
&'life0 self,
current_num_ctx: Option<usize>,
) -> Pin<Box<dyn Future<Output = Option<ModelPlacement>> + Send + 'async_trait>>where
Self: 'async_trait,
'life0: 'async_trait,
fn verify_placement<'life0, 'async_trait>(
&'life0 self,
current_num_ctx: Option<usize>,
) -> Pin<Box<dyn Future<Output = Option<ModelPlacement>> + Send + 'async_trait>>where
Self: 'async_trait,
'life0: 'async_trait,
Best-effort: where the loaded model currently sits in memory. The default
returns None (unknown / not applicable); Ollama overrides it to probe
/api/ps. Awaited only on the effect runtime, after a turn (when the
model is resident), so it never blocks the UI.
Sourcefn supports_vision<'life0, 'async_trait>(
&'life0 self,
) -> Pin<Box<dyn Future<Output = Option<bool>> + Send + 'async_trait>>where
Self: 'async_trait,
'life0: 'async_trait,
fn supports_vision<'life0, 'async_trait>(
&'life0 self,
) -> Pin<Box<dyn Future<Output = Option<bool>> + Send + 'async_trait>>where
Self: 'async_trait,
'life0: 'async_trait,
Best-effort: whether the active model can actually see images. None
means “unknown / not applicable” — the default for providers that don’t
probe, and for cloud providers whose vision support is already known
good; Some(false) is what drives the no-vision-model warning. Awaited
only on the effect runtime, so a probe never blocks the UI.
Dyn Compatibility§
This trait is dyn compatible.
In older versions of Rust, dyn compatibility was called "object safety".