raphaelsty/neural-cherche

Supports fine-tuning of dense (ColBERT) and sparse neural retrievers (Splade, SparseEmbed) on custom datasets using triplet loss, with efficient inference via pre-computed embedding caches across CPU, GPU, and MPS devices. Implements a two-stage retrieval-then-ranking architecture where BM25/sparse retrievers surface candidates that ColBERT re-ranks using token-level interactions, enabling offline embedding precomputation to avoid redundant computation.

367 stars and 1,063 monthly downloads. No commits in the last 6 months. Available on PyPI.