beir-cellar/beir

Supports multiple retrieval paradigms including lexical (BM25), dense (neural embeddings), sparse, and reranking-based approaches through a unified evaluation interface. Provides standardized implementations for diverse retrieval architectures and integrates with Hugging Face models, enabling zero-shot evaluation across heterogeneous tasks spanning different domains and query-document characteristics. Includes preprocessing pipelines for custom datasets and computes multiple metrics (NDCG, MAP, Recall, Precision, MRR) at standard cutoff points.

2,105 stars and 33,641 monthly downloads. Used by 7 other packages. Available on PyPI.