What is a semantic ID?

Definition

A catalogue can assign items short codes whose structure reflects relationships between them. Items with related content may share parts of a code, giving a model a more useful signal than an unrelated record number.

Semantic IDs can help connect new or rarely viewed items to better-known neighbors and support retrieval or ranking. A shared prefix still does not prove that two products are interchangeable or equally relevant.

ELI5

A semantic ID is a short code for an item that tries to carry useful clues about it. The code lets an AI system handle many catalogue items using a small set of symbols.

For example, a music service could give tracks codes learned from their descriptions and listening patterns. A model might then connect a request for gentle acoustic music to tracks with related codes.

How is a semantic ID different from an ordinary item ID?

An ordinary ID mainly identifies one record. A semantic ID is designed so parts of the code also reflect useful relationships between items.

Can semantic IDs help with new catalogue items?

They can help a model place a new item near related items before many people have interacted with it, though the placement still needs checking.

  1. Yves Raimond and Jacqueline Wood in blue shirts around the headline “SPOTIFY'S LLM RECS” on a black background.
  2. Devansh Tandon in a blue shirt beside the headline “LLM RECOMMENDERS” on a black background.
  3. Raghav Saboo in a blue shirt beside the headline “DISTILL THE LLM” on a black background.
Definition card for semantic ID: What is a semantic ID?

A semantic ID is a compact identifier that gives an AI system clues about an item's meaning or similarity to other items.