Lesson 2

Where Candidates Come From

Thunder, Phoenix retrieval, and SimClusters — the three sourcing systems

Three Sourcing Systems

Thunder — In-Network

thunder/ keeps recent posts from accounts you follow in memory. It's fast because it doesn't need to search — it already has the posts. Thunder has a max results cap (default: 1200) and uses a clustering algorithm to organize posts.

Phoenix — Out-of-Network Retrieval

phoenix/ is the ML retrieval system. It embeds you (based on your recent engagement) and every post as vectors, then finds the nearest neighbors. Think of it as: "what would this person likely engage with, even from accounts they don't follow?"

Phoenix has two modes:

  • Inference — uses your real-time engagement sequence
  • Retrieval — uses a separate retrieval model for broader discovery

SimClusters — Community Detection

simclusters/ groups accounts and posts into clusters based on who engages with what. If you engage with AI/ML content, you'll be associated with the AI cluster, and posts from that cluster will surface in your out-of-network candidates.

The key insight: Phoenix doesn't just find posts similar to what you've liked — it predicts what you'll do next. The embedding is forward-looking, not just a similarity match.

Check Your Understanding

1. What does Thunder store?
  1. A. All posts on X
  2. B. Recent posts from accounts you follow, in memory✓ correct
  3. C. Trending topics
  4. D. Your DM history

Why: Thunder is the in-network source — it holds recent posts from people you follow in memory for fast retrieval.

2. How does Phoenix find out-of-network posts?
  1. A. By checking what's trending
  2. B. By embedding you and posts as vectors and finding nearest neighbors✓ correct
  3. C. By looking at what your friends liked
  4. D. By random sampling

Why: Phoenix embeds the viewer and every post as vectors, then finds nearest neighbors — ML-based retrieval.

3. What does SimClusters do?
  1. A. Ranks posts by engagement
  2. B. Groups accounts/posts by engagement patterns into clusters✓ correct
  3. C. Filters spam
  4. D. Generates post recommendations

Why: SimClusters groups accounts and posts into clusters based on who engages with what — community detection.

Exercise

Think about the last 5 posts you engaged with. Were they from accounts you follow (Thunder) or accounts you don't (Phoenix/SimClusters)? This tells you which sourcing system is driving your feed.

Hint: If most are from people you don't follow, Phoenix is doing heavy lifting for you.

Where Candidates Come From | Qubax Growth Lab · Qubax AI