Named vectors are a feature that lets a single stored object carry multiple distinct vector representations, each identified by a name and typically produced by a different model or from a different aspect of the object. Rather than reducing each item to one vector, the database keeps several, and queries can target whichever is appropriate.
This is useful when one object can be meaningfully embedded in more than one way. A product might have a vector for its text description, another for its image, and another for its title; a document might have one embedding tuned for semantic search and another for a specialised task. Named vectors keep these together on the same object while allowing each to be searched independently.
The benefit is flexibility and richer retrieval without duplicating records. You can search by image similarity in one query and text similarity in another, or combine them, all against the same set of objects. Named vectors are particularly valuable for multimodal and multi-purpose applications, where a single notion of similarity is not enough to capture how the data needs to be retrieved.