Struct ModelInvocation 

pub struct ModelInvocation {
    pub model_name: String,
    pub output_column: String,
    pub feature_exprs: Vec<Expr>,
    pub feature_names: Vec<String>,
    pub feature_property_refs: Vec<(String, String)>,
    pub yield_alias: Option<String>,
    pub original_feature_exprs: Vec<Expr>,
    pub path_context: Option<PathContextFeature>,
    pub embedder_alias: Option<String>,
}

A single neural-model invocation site extracted from a clause body.

At runtime, after the clause body produces a batch of rows, each invocation evaluates its feature_exprs per row, packs them into crate::ClassifyInputs, calls the classifier in one batched classify call, then appends the result vector as a new column output_column to the batch.

Fields

model_name: String

Name of the model from CREATE MODEL <name>.

output_column: String

Synthetic column name where the per-row probabilities are written. Generated as __model_<name>_<idx> where idx is a dedup index for repeated invocations of the same model.

feature_exprs: Vec<Expr>

Argument expressions from the invocation — one per declared INPUT binding. Evaluated in clause-body scope to produce the per-row feature value passed under the binding’s variable name.

feature_names: Vec<String>

Names of the model’s INPUT bindings in declaration order, used as feature keys when building crate::ClassifyInput.

feature_property_refs: Vec<(String, String)>

Property-access expressions referenced by feature_exprs, recorded as (variable, property) pairs (e.g. for scorer(s.tier) → [("s", "tier")]). The compiler appends a matching hidden YIELD item for each so the planner’s standard property-materialization pipeline produces a column named "<variable>.<property>" in the body batch; runtime then reads from that column.

yield_alias: Option<String>

Phase C B1–B3: when the invocation appears in a YIELD item (e.g. scorer(s) AS risk), this carries the user-visible alias (risk) — distinct from the synthetic output_column (__model_scorer_0). Allows EXPLAIN to look up the model output by the column name that survives LocyProject’s projection. None when the invocation lives only inside an ALONG / FOLD expression and never surfaces as a user-visible YIELD column.

original_feature_exprs: Vec<Expr>

Phase C B1-B3 follow-up: the user-authored feature expressions BEFORE the InvocationLifter rewrote them to Variable("__model_<n>_<idx>") references. Preserved so EXPLAIN can reconstruct ClassifyInput per fact at lookup time (the rewritten feature_exprs carry synthetic-column references that can’t be evaluated against a post-projection fact_row). Same length and ordering as feature_exprs and feature_names.

path_context: Option<PathContextFeature>

Phase D D3: snapshot of the model’s path_context declaration (if any) carried onto the invocation so the runtime can pull the named column from the source rule’s derived facts at classify time without re-consulting the model catalog.

embedder_alias: Option<String>

Phase D D2 follow-up: optional embedder alias from the model’s USING xervo('classify/X', embedder='alias') clause. When None, the runtime falls back to alias "default" for semantic_match query-text embedding.

Trait Implementations

impl Clone for ModelInvocation

fn clone(&self) -> ModelInvocation

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for ModelInvocation

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.