Vertex

Reranking retrieved candidates

Retrieve broadly, then rerank precisely. A reranker reads query and passage together, which is why it beats vector distance on ordering.

Lesson 3.25m 15sAdvanced5.8k students

Chapters

5:15
  1. 0:00Background
  2. 1:29Cross Encoder
  3. 2:36Bi Encoder
  4. 3:27Differences between bi encoders and cross encoders
  5. 4:01Code Snippets for bi encoders and cross encoders
  6. 4:16Attempts to bridge the gap

What this lesson covers

  • Retrieve a wide candidate set, then narrow it
  • Score query and passage jointly with a reranker
  • Balance the latency cost against the accuracy gain

This lesson sits in Retrieval Quality, part of Retrieval-Augmented Generation from Scratch. It assumes what came before it and leads directly into the next lesson in the module.

In this lesson you will

  • Retrieve a wide candidate set, then narrow it
  • Score query and passage jointly with a reranker
  • Balance the latency cost against the accuracy gain

Resources

Previous lessonHybrid keyword and vector search3.1 · 6m 47sNext lessonEvaluating retrieval3.3 · 12m 19s