Core Recommendation Pipeline
How a post travels from the global corpus to your "For You" feed.
Candidate Sourcing
Thunder (In-Network)
Phoenix (Discovery)
Hydration
Enriching items with author data, media, and social context.
Filtering
Removing duplicates, blocks, mutes, and unsafe content.
Scoring
Phoenix Transformer
Weighted Actions
Selection
Final sorting and diversity adjustments.
Click a stage above to see technical details from the codebase.
Phoenix Transformer
The Grok-based engine that predicts your engagement.
Candidate Isolation
Unlike standard LLMs, the X algorithm uses "isolation masking" so candidates can't see each other.
User History
Candidate
Masked
Action Predictions
The model predicts 15+ different actions simultaneously.
Feed Simulator
Experiment with ranking weights to see how the feed adapts.