Reranker

AI concepts

About 1 min read

A model that precisely re-orders candidate documents retrieved from a primary search based on their relevance to the query.

Also known as

RerankerRerankingRerankingRe-ranking

Detailed explanation

A reranker is a model designed to improve search quality by precisely re-ranking candidate documents retrieved from a primary search (vector or keyword search) based on their relevance to the query. It typically uses a Cross-Encoder approach, which takes both the query and the document as inputs to perform a deep comparison; this makes it more accurate than the primary search, but slower. Therefore, it is commonly used in a two-stage system: 'narrow down candidates with a fast primary search → re-rank the top results with a reranker.' It is a key component for upgrading the quality of source documents in RAG chatbots, which is especially critical in systems where retrieval accuracy directly dictates response quality.

Why It Matters in Tool Selection

Whether a RAG or search tool supports a reranker dictates the accuracy of the retrieved sources. A primary vector search alone can return irrelevant documents at the top, which a reranker corrects. Evaluation criteria include whether the tool offers a reranking step, how well it performs with Korean documents, and whether the added latency and cost are manageable.

Key Considerations

Can you append a reranking step after the primary search?
Has the re-ranking accuracy been validated with Korean documents and queries?
Is the added latency and cost from reranking acceptable?
Can you adjust parameters, such as the number of top results to re-rank?

Real-World Examples

When an internal policy chatbot searches for 'work from home application,' the primary search quickly retrieves 20 similar documents. The reranker then promotes the top 3 documents that best fit the query's intent, and the chatbot answers based on those documents, reducing the risk of citing incorrect policies.

References

All glossary terms