A personalization vector embeds a query together with signals from a user’s past product interactions. Search first retrieves candidates, then cosine similarity against this vector reranks them to favor relevant preferences. section 9.4.3
Generate a contextual personalization vector in four steps:
- Retrieve embeddings and cluster assignments for previously interacted-with products. section 9.4.3
- Find the most similar clusters for the incoming query. section 9.4.3
- Keep only interaction vectors belonging to those query-related clusters. section 9.4.3
- Combine the query vector with the filtered interaction vectors using configurable weights. section 9.4.3
Microwave search
A user previously interacted with a Hello Kitty water bottle and a stainless-steel electric range, then searches for “microwave.” Cluster guardrails retain the range interaction, producing a contextually focused vector instead of mixing both interests.
