albertobarnabo/ecommerce-product-search-embeddings
Sentence Similarity • 33.4M • Updated • 264 • 1
Self-hosted product search on CPU: embeddings + reranker + data, trained on 427k real Amazon judgments. No per-query fees.
Note The flagship: 33M retriever, nDCG@10 0.7483 vs 0.7250 base. Start here.
Note Stage 2: +8.1 nDCG end-to-end reranking the retriever's top-50, 45 ms/query on CPU.
Note Quality variant: 109M, +2.1 Recall@100 on a 1.22M-product corpus.
Note 31MB model2vec static: 73k queries/s on CPU, for edge and serverless.
Note The training pairs: real ESCI judgments + provenance-labeled synthetic noise.