vintasoftware/entity-embed

Combines PyTorch Lightning for training with the N2 approximate nearest neighbor library to enable efficient blocking at scale, achieving ~0.99 recall with minimal false negatives. Uses contrastive learning to organize embeddings in N-dimensional space where duplicate records cluster together, optimized specifically for the blocking/indexing stage rather than final pair matching. Designed as a preprocessing component in entity resolution pipelines, returning high-recall candidate pairs for downstream filtering by pairwise classifiers.

161 stars. No commits in the last 6 months. Available on PyPI.