OneSKU

Overview

OneSKU is a hybrid retrieval system for product search across heterogeneous vendor catalogs, where the same item appears under different names, units, and attribute schemas per vendor.

Impact: sub-15 second query latency across multi-million SKU inventories.

How it works

• Hybrid retrieval: BM25 lexical matching fused with BERT-based dense embeddings, so exact part numbers and fuzzy descriptions both work.
• Harmonization pipeline: custom normalization that aligns vendor catalogs (units, attribute names, category taxonomies) before indexing.
• Built in PyTorch with offline evaluation on labeled query-product pairs.

What I'd highlight

Hybrid wasn't a buzzword choice: pure dense retrieval failed on exact SKU lookups and pure BM25 failed on descriptive queries. The fusion weighting was tuned against real query logs.