Member Junction
    Preparing search index...

    Interface VisionLLMFeatureStep

    vision-llm — multimodal vision-LLM-as-feature (plan §11 / §5.6, PS-MM-1). For each row, a vision prompt is run over the row's image (URL or binary ref) and its single structured output (a named category or scalar) becomes the feature column's value.

    Unlike preprocessing steps (impute/standardize/onehot/bin), this step is a per-row, stateless extraction: there is no fitted state — the same prompt is applied to each row independently — so it is exempt from the fit-once/apply-everywhere split and produces a RAW column directly, exactly like select/embedding (see the executor for the full reasoning).

    interface VisionLLMFeatureStep {
        Id: string;
        ImageColumn: string;
        Inputs?: string[];
        Kind: "vision-llm";
        Label?: string;
        ModelRef?: string;
        Output: VisionLLMOutput;
        Prompt: VisionPromptRef;
    }

    Hierarchy (View Summary)

    Index

    Properties

    Id ImageColumn Inputs? Kind Label? ModelRef? Output Prompt

    Properties

    Id: string

    Stable id for this step, referenced by downstream steps' Inputs.

    ImageColumn: string

    The row field holding the image to analyze — either a fully-qualified image URL or a binary/data reference (e.g. a data:image/...;base64,... data URL). Resolved per-row at assembly time; rows with a null/missing value yield a null feature (no prompt is run).

    Inputs?: string[]

    Ids of upstream steps that feed this step (DAG edges).

    Kind: "vision-llm"

    Discriminator selecting the concrete step variant.

    Label?: string

    Optional human-readable label for the visual DAG.

    ModelRef?: string

    Optional model override (an MJ: AI Models id/name) forcing a specific vision-capable model. When omitted, the prompt's own model selection applies. Pinned into the model's Lineage so scoring uses the same vision model.

    Output: VisionLLMOutput

    The expected structured output the model emits — the named scalar/category that becomes this step's feature column.

    Prompt: VisionPromptRef

    The vision prompt to run over the image. Either a reference to a stored MJ: AI Prompts row (id or name) or an inline prompt body. Exactly one form is used; when both are set the inline body takes precedence.

    Settings

    Member Visibility

    On This Page

    Properties