use case · wild
A hard negative is a passage stuffed with the query's exact words that nonetheless answers the wrong question — the kind of thing keyword search ranks first and bi-encoder embeddings often can't tell apart. Because a cross-encoder reads the query and passage together, it can catch the mismatch. The contrast below ranks by the cross-encoder on the left and by shared-word overlap on the right, so you can watch the trap passages fall.
The lexical column is exactly what a naive keyword engine would return — the decoys, packed with the query's words, float to the top. The cross-encoder reads for meaning and demotes them. This is why reranking is worth a second, heavier pass: it fixes precisely the mistakes cheap retrieval makes.
← Practical · Back to BGE-reranker · Next: Multi-model →