02.04 · Lecture
Rerankers: How a Passage Becomes a Citation
Explain what a reranker does that first-stage retrieval cannot, and describe the path from candidate document to the specific passage a model cites.
A reranker is the slower relevance judge after initial retrieval, comparing the user query against candidate passages closely enough to choose evidence, not just nearby documents. It turns a broad shortlist into ordered, answer-shaped chunks that can be placed in context and later shown as citations.
What this lesson answers
- what does a reranker do in RAG
- why is first stage retrieval not enough
- how does a document become an AI citation
Notes
A search system usually works in stages. The first stage retrieval step is built to be fast: given a user question, it pulls back a rough candidate set from a large corpus using keyword matching, vector similarity, or both. That first pass is good at recall, meaning it tries not to miss potentially useful documents, but it is often too crude to decide which exact passage is most relevant to the user’s specific need.
A reranker is the slower, more careful second judge. Instead of asking “which documents are near this query in an index?”, it compares the query with each candidate passage more…
Common questions
- What does a reranker add after vector or keyword retrieval?
- First-stage retrieval is optimised to find plausible candidates quickly across a large corpus. A reranker spends more work on a smaller set, comparing the query and each passage more directly. That helps separate a passage that actually answers the question from one that merely shares words, entities, or general semantic neighbourhood.
- How does a source document become a cited passage?
- The document is split into chunks, then retrieval selects candidate chunks for the query. A reranker reorders or filters those chunks by relevance. The strongest passages are added to the model context with source metadata. The application can then attach a citation back to the passage used as supplied evidence.
- Does a citation prove the model used that source correctly?
- No. In many RAG systems, a citation means the passage was retrieved, ranked highly, and supplied to the model. It does not by itself prove every generated sentence is supported by that passage. Reliable grounding needs additional checks that map claims back to the evidence.
Short definition: what is Rerankers: How a Passage Becomes a Citation?