How Modern Ranking Systems Work (www.shaped.ai)

This piece outlines a practical, component-based view of modern multi-stage ranking systems—what used to be taught as a linear funnel (Retrieval → Filtering → Scoring → Ordering) is reframed as a Directed Acyclic Graph of interacting components. Using a "perfect_score(user,item,context)" thought experiment, the article explains why real systems approximate an impossible oracle under tight latency (e.g., P99 ≈ 200 ms) and cost constraints, forcing engineering trade-offs between relevance, latency, and cost. It traces the design’s evolution through canonical posts and papers, then distills core patterns: parallel candidate sources (vector search, collaborative filtering, popularity heuristics), specialized rankers for ads or cold-start items, multi-point filtering, and the final listwise ordering step. Technically salient points and practical cautions are emphasized: retrieval produces ~1k candidates for expensive scoring (DLRM and increasingly Transformer-based sequential models), filtering should shrink search space early, and production systems must avoid “hero retriever” single-source blind spots and offline/online feature skew. The author frames the work as a series that will dig into each stage—next up, retrieval—promising code for a collaborative-filtering model and guidance on ANN index choices, with concrete trade-offs between recall, latency, and cost that drive real-world design decisions.