AgentSkillsCN

x-retrieval-systems

在分析系统如何将每日 5 亿条推文筛选为约 1,500 条候选推文时，可运用此技能。这是“漏斗顶端”的工程设计，决定了哪些内容有资格进入排名环节。

SKILL.md
--- frontmatter
name: x-retrieval-systems
description: Use this skill when analyzing how the system narrows down 500 million daily tweets to a candidate pool of ~1,500. This is the "top-of-funnel" engineering that determines what is even eligible to be ranked.
version: 1.0.0
license: MIT

X Retrieval Systems

Expertise in X's multi-stage retrieval architecture, including Earlybird search indexing and Phoenix-based ANN similarity search.

Context

Retrieval at X is split between In-Network (content from people you follow) and Out-of-Network (discovery). In-Network retrieval relies on Earlybird (Lucene-based search), while Out-of-Network retrieval uses Phoenix (Two-Tower embeddings) and ANN (Approximate Nearest Neighbor) algorithms like HNSW.

What it does

  • •Decodes In-Network Sourcing: Explains how Earlybird shards the index into Realtime, Protected, and Archive clusters.
  • •Explains Discovery Logic: Details how Two-Tower models enable "semantic" search for content you don't follow.
  • •Analyzes Latency: Breaks down the single-writer/multi-reader concurrency model that allows for sub-second global retrieval.

Example Trigger Prompts

  • •"/find-candidates how Earlybird shards real-time index"
  • •"/find-candidates retrieving 1,500 candidates from 500M tweets"
  • •"Role of HNSW in embedding-based discovery"
  • •"In-Network (Thunder) vs Out-of-Network (Phoenix) retrieval"
  • •"Trace 'Discovery' request: User Embedding → Candidate Source"