SYSTEM FLOW

01

Query Hydration

Fetch user context: engagement history, following list, preferences.

02

Candidate Sources

Thunder (followed accounts) + Phoenix Retrieval (global tweets).

03

Hydration

Enrich with tweet content, author info, media metadata.

04

Filters

Remove duplicates, old posts, self-posts, seen posts, blocked/muted authors.

05

Scoring

Phoenix predicts multi-action probs → weighted sum → diversity decay.

06

Selection + VF

Take Top-K → visibility filtering → final distribution.

Interactive Pipeline

Explore each stage in detail with our interactive explorer tool.

Ranking Signals

Deep dive into the Attention Mask and candidate isolation logic.

CORE MECHANICS

PREDICTION

System predicts P(like), P(reply), P(repost) simultaneously using multi-target Transformers.

DIVERSITY

Same-author posts are attenuated via Author Diversity heuristics to avoid feed homogenization.

ISOLATION

Candidates are scored independently via Attention Masking for stability and efficient caching.

NETWORK

Out-of-network content is decayed based on the Social Graph relationship strength.