Phoenix

"ML component with two functions: (1) two-tower retrieval for out-of-network post discovery, and (2) Grok-based transformer ranking that predicts engagement probabilities."

Phoenix
├── Retrieval   →  find relevant posts from global corpus
└── Ranking     →  score each candidate post

Architecture

User engagement history          All posts in corpus
         │                               │
         ▼                               ▼
    User tower                      Post tower
  (hash embeddings)              (hash embeddings)
         │                               │
         ▼                               ▼
   User embedding              Post embeddings (all)
         │                               │
         └──────────┬────────────────────┘
                    ▼
           Dot product similarity
           user · post = cos(θ)   [after L2 normalisation]
                    │
                    ▼
              Top-K candidates  →  out-of-network pool

How dot product similarity works

Both vectors are L2-normalised, so the dot product equals the cosine of the angle between them:

$$\text{similarity} = \mathbf{u} \cdot \mathbf{p} = \cos(\theta)$$

Posts closest to the user vector in embedding space become out-of-network candidates.

Hash-based embeddings

Rather than a fixed vocabulary lookup table, both towers use multiple hash functions per feature. This handles an unbounded feature space (billions of post IDs, author IDs, etc.) without a rigid vocab.

Input

Candidate isolation

"Posts cannot attend to each other during inference. Each post only attends to the user context."

This is a deliberate design decision:

Output — action probabilities

The model outputs ~15 probabilities per post:

Scoring formula

$$\text{Score} = \sum_{i} w_i \cdot P(\text{action}_i)$$

Positive actions get positive weights; negative actions get negative weights, actively suppressing unwanted content.