False entity merges don’t just dirty data. They distort inventory, pricing, and forecasts, then every model and report built on top. Learned sparse retrieval improves recall, but it can still treat records like unordered tokens. This post adds token-to-token attention as a structural check so near-duplicates pass and lookalikes fail, with a trail you can audit.

Entity resolution breaks when exact matching is too brittle and dense vectors blur identities. This post introduces SPLADE, a learned sparse retrieval model that keeps inverted indexes and token-level explainability while adding transformer-powered expansion and reweighting. We walk through where SPLADE beats BM25 and dense search, where it can fail on SKUs and over-expansion, and how to run it in Postgres/ParadeDB for large-scale product, customer, or patient identity.