← All models · MS MARCO reranker
use case · practical
Embedding search is fast but blunt — it compares two vectors made independently, so it can rank a passage high just for sharing a topic. A cross-encoder reads the query and passage together and is far more precise, but too slow to score a whole corpus. The production answer is both: embeddings shortlist, the reranker reorders. Below, watch the same results ranked by embedding cosine and then by the reranker — the column that puts the real answer first is the reranker.
Loads two models (~25 MB embeddings + ~90 MB reranker). Both auto-initialise when cached.
Stage 1 · MiniLM embeddings
Stage 2 · MS MARCO reranker
Watch the "Forgot password?" reset-email line: embeddings often rank the generic "change your password" instruction first because it shares more words, but for the intent "I'm locked out" the reranker lifts the reset-email passage — because it actually answers the question.
← Basics · MS MARCO reranker · Next: Wild →