Engineer breaks down X's search ranking stack: two-tower retrieval plus L1/L2 rankers
_jaydeepkarale · x · 2026-09-16
A developer walks through the layered architecture of (presumably) X's search/recommendation system:
Retrieval layer
- Token-based retriever: fetches posts matching keywords
- Embedding-based retriever: a two-tower architecture converts query and posts into embeddings via AI models, ranked by cosine similarity
Multi-stage ranking layer
- L1 and L2 rankers share a similar architecture, differing in size and parameters passed
- Each stage outputs on-topicness and long-dwell scores
- These scores determine which posts are finally shown to users
More from Infra
- Kastrup's Memory-in-Compute Chip Startup Raises €200M+ With Samsung Co-Leading — demian_ai · 2026-09-17
- Cloud overlays saturated his network, so he runs AI camera analytics on a local edge appliance — Chris__Codes · 2026-09-17
- Google open-sources Agent Substrate: millions of agent sandboxes at 10x density on GKE — rseroter · 2026-09-17
- Chipotle partners with Palantir on Foundry-based food safety risk platform — eliano · 2026-09-17
- Emerald AI, Google and NVIDIA launch alliance for flexible AI data centers — ArtificialOther · 2026-09-17
- Corv v1.1.1 ships as an AI-native SSH execution layer purpose-built for infra agents — Dude01_ · 2026-09-17