Pipeline Explorer
Deconstruct the For You recommendation flow. Below are the complete stages of the recommendation engine from request to delivery.
Query Hydration
Gathers user context: recent actions, following list, and preferences. The system builds a "who are you" profile for this specific request.
Maintain a clear profile niche so the system can hydrate accurate context for your target audience.
Candidate Retrieval
Retrieves ~1,500 initial candidates from In-Network (Followers) and Out-of-Network (Social Graph/Embeddings) sources.
Follower engagement is the baseline; Embeddings are your ticket to reaching non-followers.
Filtering & Safety
Excludes duplicates, stale content, blocked authors, or muted keywords. Visibility Filtering (VF) rules are applied here.
Avoid reposting identical content to stay clear of deduplication filters. Maintain high account health.
Heavy Ranking
The core stage. Uses Transformer models (Phoenix) to predict interaction probabilities (Likes, Replies, RTs) for each tweet.
Design for high-value signals: replies and dwell time have massive impact on the weighted sum.
Selection & VF
Selects Top-K and performs final safety, diversity (author/topic), and fatigue checks before delivery.
Avoid spamming multiple posts at once to prevent Author Diversity attenuation.