Re-ranking
Re-ranking reassesses search candidates already retrieved. Relevance to the query and the limits of the initial selection both matter.
Ordering retrieved candidates more carefully
Re-ranking means reassessing and ordering search results already retrieved. An initial search selects possible documents or passages. A subsequent method examines their relevance to the particular query more closely, for example with a cross-encoder that processes the query and text together. This step cannot discover documents outside that selection. Access permissions and business filters must also apply to the reordered results.
A fictional service portal receives a question about how customers can check the repair status of a device they sent in. The initial selection contains a general contact page and instructions for checking status. A re-ranker could move the relevant instructions ahead of the contact page, even if that page initially ranked higher. A planned review would check whether the right passage is present among the candidates and then ranked appropriately. The instructions explain how to make an enquiry; their ranking does not provide the customer's actual repair status.
Sources
Related terms
- Semantic searchSearch that finds and ranks results by their semantic relevance to the query.
- Hybrid searchCombining lexical search and vector search into a shared result list.
- Document chunkingSplitting a document into smaller coherent sections for search and model context.
- Retrieval-augmented generationRAG combines information retrieval with generated answers. Access to sources does not remove the need for quality and permission design.
- LLM evaluationLLM evaluation tests model outputs against defined tasks and criteria. One good example does not establish overall quality.
- Model routingChoosing a suitable model for each request from a defined set of permitted candidates.
Your project with ASES
Describe your project to us. We will get back to you with an initial assessment.