Reranking reassesses the order of an existing candidate set for a question. It normally complements initial retrieval rather than searching an entire collection by itself.
Retrieve, then compare
Initial retrieval quickly selects plausible items from a larger scope. A later method can compare them more carefully. Cross-Encoders are one public pair-scoring approach, not a requirement for every reranker.
Rankings can be wrong
A higher item may be more relevant without being newer, more reliable or sufficient for every claim. Reranking cannot recover material absent from the initial candidates.
Helping RAG
Better candidates may improve the opportunity to use material. Evaluation still needs citation checks, conditions and generated results, beyond a reranker's isolated score.
Is the first result the correct answer? No. Relevance ranking and fact-checking are different work.
Continue with What is Hybrid Search? Keywords and vectors to compare related ideas.
For a related project perspective, explore LingoAI’s AI work.