Course Wiki

Reranking

A second-pass model that reorders retrieved documents by true relevance to the query.

Reranking is precision work on top of fast retrieval: retrieve a generous candidate set (say 50 documents) with cheap vector or keyword search, then score each candidate against the query with a heavier cross-encoder model that actually reads query and document together, and keep only the top few for the prompt.

The course positions it as the last quality lever before generation: first-stage retrieval optimizes recall at speed, the reranker optimizes the ordering of what reaches the LLM. Better context ordering measurably improves answers — which the evaluation module’s harness is there to verify.

Taught in

• LLM Zoomcamp

  • Module 2: Vector Search
    • Vector Search with sqlitesearch
    • Next Steps
  • Module 6: Best Practices
    • Best Practices for RAG
    • Document Reranking
    • Next Steps ## Related concepts

• Hybrid Search

• Vector Search

• LLM Evaluation

• RAG

• Keyword Search