LLM Learning Hub

workspace/llm-course/home

Reranking

Reranking in vector search re-scores the top retrieved candidates with a cross-encoder that jointly encodes query and document, producing a more accurate relevance ordering than bi-encoder similarity.

What is Reranking?

Reranking in vector search re-scores the top retrieved candidates with a cross-encoder that jointly encodes query and document, producing a more accurate relevance ordering than bi-encoder similarity.

Reranking in vector search re-scores the top retrieved candidates with a cross-encoder that jointly encodes query and document, producing a more accurate relevance ordering than bi-encoder similarity.

Where is it used?

Cohere Rerank API, `BAAI/bge-reranker-v2-m3`, and `sentence-transformers` cross-encoders are applied after HNSW or IVF search in LlamaIndex and Haystack pipelines.

How to build it

Retrieve top-50 with dense search, load `CrossEncoder('BAAI/bge-reranker-base')`, call `model.rank(query, docs, top_k=5)`, and return the top-5 re-scored documents for generation.