Amazon's library for when your label space is absurdly large

What it does PECOS is Amazon’s framework for extreme multi-label ranking and classification — problems where the output space is comically large (think 100 million labels). It trains hierarchical models that recursively narrow down candidates, then ranks the survivors. The core C++ inference path is built for real-time serving, and there’s a Transformer-based variant (XR-Transformer) that fine-tunes Hugging Face models for the same task.

The interesting bit The hierarchy trick is the whole game: instead of scoring every label, PECOS builds a label tree and learns linear models that traverse from root to promising leaf clusters. It’s old-school divide-and-conquer dressed up for modern scale. The ANN module (HNSW with SIMD optimizations) is a nice bonus for dense/sparse retrieval, though it’s presented as a supporting player rather than the main event.

Key highlights

• X-Linear: recursive linear models with C++ inference, handles 100MM output space
• XR-Transformer: SotA on public XMC benchmarks, plugs into Hugging Face ecosystem
• HNSW ANN search with SIMD-optimized distance computation for sparse and dense features
• Thread-safe parallel graph construction and inference searchers
• Apache 2.0 licensed, pip-installable as libpecos

Caveats

• Platform support is narrow: Ubuntu 20.04/22.04 and Amazon Linux 2 only
• The “general-purpose framework” claim is aspirational — the README examples and papers are almost entirely XMC/text classification flavored

Verdict Worth a look if you’re serving extreme multi-label models at scale and need the C++ speedup. Skip it if your label space is merely large (thousands, not millions) — you probably don’t need the complexity.