Trait VlmEngine 

pub trait VlmEngine: Send + Sync {
    // Required methods
    fn info(&self) -> EngineInfo;
    fn describe(
        &self,
        prompt: &VlmPrompt<'_>,
        opts: &VlmOptions,
    ) -> Result<String>;
    fn describe_video(
        &self,
        frames: &[VideoFrame],
        opts: &VlmOptions,
    ) -> Result<Vec<TimedSegment<String>>>;

    // Provided method
    fn describe_image(
        &self,
        image: &ImageBuffer,
        opts: &VlmOptions,
    ) -> Result<String> { ... }
}

Image/video → description (Argus).

See VlmOptions for the Gate-2 surface decision and the written-down v1 exclusions.

Required Methods

fn info(&self) -> EngineInfo

fn describe(&self, prompt: &VlmPrompt<'_>, opts: &VlmOptions) -> Result<String>

The general path: an ordered, interleaved multimodal prompt.

This is the required method, and describe_image is derived from it, rather than the other way round. An engine that implemented only the single-image case would still compile against a multi-image prompt — and would then silently answer using the first image, or the last, or a concatenation. Making the general case the one an implementor must write means multi-image support is a compile-time obligation instead of a runtime surprise.

fn describe_video( &self, frames: &[VideoFrame], opts: &VlmOptions, ) -> Result<Vec<TimedSegment<String>>>

Video understanding over sampled frames → a timed caption track.

Frames arrive with their timestamps (VideoFrame::timestamp) so an engine can encode time and not merely order — M-RoPE’s third axis. Whether it does is the engine’s business; the trait’s job is to make sure the information reaches it.

Provided Methods

fn describe_image( &self, image: &ImageBuffer, opts: &VlmOptions, ) -> Result<String>

One image, with VlmOptions::prompt as the instruction.

Provided: builds a single-image VlmPrompt and calls Self::describe. Zero-copy — the prompt borrows the image.

Dyn Compatibility

This trait is dyn compatible.

In older versions of Rust, dyn compatibility was called "object safety".

Implementors