LEVEL 02 · LESSON 18
Reranking
Retrieval gathers candidates; reranking prioritizes the evidence to read.
What it is
Reranking is a second pass that reorders retrieved candidates using a more focused assessment of their relevance to the question.
Why it matters
Fast retrieval often casts a wide net. Reranking helps prioritize the passages that will be most useful within a limited context budget.
A business example
A delayed-delivery question retrieves general returns, shipping, and refund pages. A reranker places the policy specifically covering delayed shipments above broadly related documents.
How it actually works
Retrieve a candidate set, score candidates with a reranking model or method, and pass the strongest results to generation. Some rerankers consider the query and passage together, capturing relationships that independent embeddings miss. More candidates can improve coverage but increase latency and cost. Measure whether the right evidence moves upward, and remember that reranking cannot recover a document absent from the candidate set.
A common misconception
Reranking is not fact-checking. A highly relevant passage can still be outdated or wrong.
ONE SENTENCE TO REMEMBER
Retrieval gathers candidates; reranking prioritizes the evidence to read.
CHECK YOUR UNDERSTANDING
After adding a reranker, a support assistant still sometimes misses the right policy document entirely. What is a likely explanation?
Show the answer
Answer: C.
Reranking reorders retrieved candidates using a more focused relevance assessment. It cannot recover a document absent from the candidate set, and it is not fact-checking: a highly relevant passage can still be outdated or wrong.
Take it into your next meeting
Are your best sources missing entirely or just buried below weaker matches?
One more concept, understood.
Mark this lesson complete to keep your place.