← All models · MS MARCO reranker
use case · multi-model
This is the pattern behind modern search and RAG. MiniLM embeddings score the whole corpus cheaply and hand the top-N candidates to this cross-encoder, which re-reads each one against the query and produces the final, precise ordering. Fast net, then sharp filter — and both models run locally, so the entire pipeline is on-device.
Loads two models (~25 MB embeddings + ~90 MB reranker). Both auto-initialise when cached.
Stage 1 · MiniLM embeddings (retrieve)
Stage 2 · MS MARCO reranker (rerank)
The win is efficiency: the cross-encoder only ever scores the shortlist, not the whole corpus, so you get cross-encoder precision at retrieval-scale cost. Scale the corpus up and the two-stage split is what keeps it fast.