Definition
SPLADE (Sparse Lexical and Expansion) is a neural information retrieval model that transforms text into high-dimensional sparse vectors by predicting relevant vocabulary terms, even those not present in the original text. In RAG pipelines, it bridges the gap between keyword matching and semantic search, though it requires more storage and inference compute than traditional BM25.
Distinguish from dense embeddings; SPLADE produces sparse, interpretable vectors mapped to a fixed vocabulary rather than low-dimensional continuous latent space.
"An automated librarian who reads a book and scribbles relevant synonyms and related topics in the margins so the book can be found even if the reader uses different words than the author."
- Sparse Vector(Component)
- BM25(Prerequisite)
- Hybrid Search(Component)
- Vocabulary Mismatch(Problem Solved)
Conceptual Overview
SPLADE (Sparse Lexical and Expansion) is a neural information retrieval model that transforms text into high-dimensional sparse vectors by predicting relevant vocabulary terms, even those not present in the original text. In RAG pipelines, it bridges the gap between keyword matching and semantic search, though it requires more storage and inference compute than traditional BM25.
Disambiguation
Distinguish from dense embeddings; SPLADE produces sparse, interpretable vectors mapped to a fixed vocabulary rather than low-dimensional continuous latent space.
Visual Analog
An automated librarian who reads a book and scribbles relevant synonyms and related topics in the margins so the book can be found even if the reader uses different words than the author.