Pipeline Visualization

📱

1. Feed Request

User opens For You tab

▼

User ID Extracted from session

Features Loaded Interests, Following, History

Request Limit ~100 posts

⚡

2. Thunder (In-Network)

Posts from accounts you follow

▼

Source Following Timeline

Candidates ~500 posts

Max Age 7 days

Weight Factor 1.0× (full weight)

thunder/realtime_query.rs
→ Get recent posts from followed users
→ Apply freshness filter (7 day max)
→ Return candidate pool

🔥

3. Phoenix (Out-of-Network)

ML-based content discovery

▼

Source SimClusters + Social Graph

Candidates ~1000 posts

Model Grok Transformer

Weight Factor 0.7× (OON discount)

phoenix/ranker.py (JAX)
→ SimClusters user embedding
→ Grok engagement prediction
→ Return discovery candidates

🚫

4. Filtering

Remove spam, blocks, duplicates

▼

Input ~1500 candidates

Spam Filtered ~200 removed

Blocked/Muted User-specific removal

Output ~1200 candidates

⚖️

5. Weighted Scoring

Apply algorithm weights

▼

Reply Weight ×27.0

Profile Click ×12.0

Bookmark ×4.0

Report Penalty ×-369.0

home-mixer/scorers/weighted_scorer.rs
Score = Σ (weight × P(action))
+ SIMD optimization
+ Author diversity decay

✅

6. Selection & Serving