Enrich

Enrich items with ML-computed features based on user profile and preferences.

This endpoint computes personalized features for a list of items by comparing them against a user's profile. The computed features are designed to be used as inputs to the Ranking pipeline.

Semantic Features:

• Dense vector similarity between user and item embeddings (cosine similarity)
• Multiple embedding comparisons: sem_sim_1 through sem_sim_5 for primary through fifth user embeddings
• sem_sim_fuzzy: Loose matching across all embeddings
• sem_sim_closest: Maximum similarity across all user embeddings
• Cluster indicators (sem_sim_cluster1–sem_sim_cluster5): Binary flags indicating which user embedding is the closest match

Topic Features:

• usr_primary_labels / usr_primary_tags: Count of matching labels/tags against user's primary preferences
• usr_secondary_labels / usr_secondary_tags: Count against secondary preferences
• usr_mixed_labels / usr_mixed_tags: Weighted blend (75% primary + 25% secondary)
• AI:* / TAG:*: One-hot encoded features for individual labels and tags

Composite Features:

• topic_similarity: Combined score aggregating semantic and topic signals
• original_rank: Normalized position from request order (0–1, higher = earlier)

Index-Specific Features (Polymarket):

• num_bets: Number of bets on the market
• num_price_changes: Number of price change events

All numerical features are normalized to [0, 1] range for consistent scaling.