SPLADE: Sparse Lexical and Expansion Model for First Stage Ranking

SPLADE (Sparse Lexical and Expansion model), published by Thibault Formal, Benjamin Piwowarski, and Stephane Clinchant in July 2021, sits between two retrieval traditions. Classic search engines use sparse, term-based representations like BM25 that match exact words through an inverted index; modern neural retrievers use dense embeddings that capture meaning but lose exact-match precision. SPLADE produces representations that are sparse, like the classic approach, but learned by a neural network, so it can also expand a query or document with related terms the original text did not contain.

The model uses explicit sparsity regularization and a log-saturation effect on term weights to keep the output sparse and controllable, and it is trained end-to-end in a single stage. Because the result is sparse and term-based, it can be served with the same efficient inverted-index machinery that powers traditional search, while still benefiting from neural semantic learning.

SPLADE became a popular choice for the first-stage ranking step in retrieval pipelines, where you need to pull a candidate set from millions of documents quickly. For a business, it offers a middle path: the speed and exact-match reliability of keyword search combined with the semantic recall of neural models, often without giving up existing search infrastructure.

Sources

Last verified June 7, 2026