splade

Why we included this project

SPLADE comes from NAVER Labs Europe and is the reference implementation of a sparse neural retrieval model. Dense retrievers map text into fixed vectors that are hard to inspect; SPLADE instead expands each query and document into sparse, weighted term lists. Because those lists are sparse, retrieval still runs on an inverted index, and because they are terms, the matches stay interpretable, which is a real advantage when you need to explain why a result ranked. The repo covers training, indexing, and query-time retrieval, and it ships with a BEIR evaluation harness plus pre-trained v2 weights on Hugging Face, so you can test the released models before fine-tuning your own. It reads as a research codebase more than a ready-to-deploy service, but for anyone building first-stage ranking and wanting lexical transparency with neural generalization, it's the most complete place to start.