Module embed_caps 

Embedding-capability mapping (task→heads) and per-alias head requirements, shared by open-time validation and write-time auto-embed routing. Embedding-capability model: which vector heads an alias’s task can produce, and which heads a schema’s auto-embed columns require.

Auto-embed always sources from a text column, so an alias is usable for a column only if its Uni-Xervo ModelTask can produce — from text — the head that column needs (dense / sparse / multi-vector). This module is the single source of truth for that mapping, shared by open-time schema validation in the uni-db crate and by write-time routing in crate::runtime::writer.

The task→heads mapping is a deliberate allow-list. ModelTask is #[non_exhaustive], so any task that is not explicitly an embed-from-text task (image/audio/multimodal embedders, rerankers, generators, …) maps to no heads and is rejected, rather than silently slipping through when a future variant is added upstream. The mapping returns the task’s upper bound; a specific hybrid model may expose fewer heads (its available_heads() is the runtime ground truth), which the writer enforces at inference time.

Structs

RequiredHeads

Embedding heads a single alias must produce, with the columns that require them.

Functions

required_embed_heads

Per-alias embedding-head requirements implied by a schema’s auto-embed indexes.

text_embedding_heads

Vector heads an alias’s task can produce from a text source column.